METHODS AND APPARATUS TO GENERATE COMPUTER-TRAINED MACHINE LEARNING MODELS TO CORRECT COMPUTER-GENERATED ERRORS IN AUDIENCE DATA

FIELD OF THE DISCLOSURE

This disclosure relates generally to monitoring audiences, and, more particularly, to methods and apparatus to generate computer-trained machine learning models to correct computer-generated errors in audience data.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram illustrating an example system to enable the generation of audience measurement metrics based on the merging of data collected by a database proprietor and an audience measurement entity (AME).

FIG. 2 is an example block diagram of the example model generator of FIG. 1.

FIG. 3 is an example block diagram of the example model analyzer of FIG. 1.

FIG. 4 is a flowchart representative of example machine readable instructions which may be executed to implement the example model generator of FIGS. 1 and/or 2 to generate computer-generated machine learning models and associated performance results.

FIG. 5 is a flowchart representative of example machine readable instructions which may be executed to implement the example model analyzer of FIGS. 1 and/or 3 to aggregate the performance results and select one or more of the computer-generated machine learning models to use in correcting computer-generated errors in audience data.

FIG. 6 is a block diagram of an example processing platform structured to execute the instructions of FIGS. 4 and/or 5 to implement the example model generator of FIGS. 1 and/or 2 and the example model analyzer of FIGS. 1 and/or 3 to generate a plurality of computer-generated machine learning models and select one or more of the computer-generated machine learning models based on performance data to correct computer-generated errors in audience data.

The figures are not to scale. In general, the same reference numbers will be used throughout the drawing(s) and accompanying written description to refer to the same or like parts.

Unless specifically stated otherwise, descriptors such as “first,” “second,” “third,” etc. are used herein without imputing or otherwise indicating any meaning of priority, physical order, arrangement in a list, and/or ordering in any way, but are merely used as labels and/or arbitrary names to distinguish elements for ease of understanding the disclosed examples. In some examples, the descriptor “first” may be used to refer to an element in the detailed description, while the same element may be referred to in a claim with a different descriptor such as “second” or “third.” In such instances, it should be understood that such descriptors are used merely for identifying those elements distinctly that might, for example, otherwise share a same name. As used herein, “approximately” and “about” refer to dimensions that may not be exact due to manufacturing tolerances and/or other real world imperfections. As used herein “substantially real time” refers to occurrence in a near instantaneous manner recognizing there may be real world delays for computing time, transmission, etc. Thus, unless otherwise specified, “substantially real time” refers to real time+/−1 second.

DETAILED DESCRIPTION

Audience measurement entities (AMEs) usually collect large amounts of audience measurement information from their panelists including the number of unique audience members for particular media and the number of impressions corresponding to each of the audience members. Unique audience size, as used herein, refers to the total number of unique people (e.g., non-duplicate people) who had an impression of (e.g., were exposed to) a particular media item, without counting duplicate audience members. As used herein, an impression is defined to be an event in which a home or individual accesses and/or is exposed to media (e.g., an advertisement, content, a group of advertisements and/or a collection of content). Impression count, as used herein, refers to the number of times audience members are exposed to a particular media item. The unique audience size associated with a particular media item will always be equal to or less than the number of impressions associated with the media item because, while all audience members by definition have at least one impression of the media, an individual audience member may have more than one impression. That is, the unique audience size is equal to the impression count only when every audience member was exposed to the media only a single time (i.e., the number of audience members equals the number of impressions). Where at least one audience member is exposed to the media multiple times, the unique audience size will be less than the total impression count because multiple impressions will be associated with individual audience members. Thus, unique audience size refers to the number of unique people in an audience (without double counting any person) exposed to media for which audience metrics are being generated. Unique audience size may also be referred to as unique audience, deduplicated audience size, deduplicated audience, or audience.

Techniques for monitoring user access to an Internet-accessible media, such as digital television (DTV) media and digital content ratings (DCR) media, have evolved significantly over the years. Internet-accessible media is also known as digital media. In the past, such monitoring was done primarily through server logs. In particular, media providers serving media on the Internet would log the number of requests received for their media at their servers. Basing Internet usage research on server logs is problematic for several reasons. For example, server logs can be tampered with either directly or via zombie programs, which repeatedly request media from the server to increase the server log counts. Also, media is sometimes retrieved once, cached locally and then repeatedly accessed from the local cache without involving the server. Server logs cannot track such repeat views of cached media. Thus, server logs are susceptible to both over-counting and under-counting errors.

As Internet technology advanced, the limitations of server logs were overcome through methodologies in which the Internet media to be tracked was tagged with monitoring instructions. In particular, monitoring instructions (also known as a media impression request or a beacon request) are associated with the hypertext markup language (HTML) of the media to be tracked. When a client requests the media, both the media and the impression request are downloaded to the client. The impression requests are, thus, executed whenever the media is accessed, be it from a server or from a cache.

The beacon instructions cause monitoring data reflecting information about the access to the media (e.g., the occurrence of a media impression) to be sent from the client that downloaded the media to a monitoring server. Typically, the monitoring server is owned and/or operated by an AME (e.g., any party interested in measuring or tracking audience exposures to advertisements, media, and/or any other media) that did not provide the media to the client and who is a trusted third party for providing accurate usage statistics (e.g., The Nielsen Company, LLC). Advantageously, because the beaconing instructions are associated with the media and executed by the client browser whenever the media is accessed, the monitoring information is provided to the AME irrespective of whether the client is associated with a panelist of the AME. In this manner, the AME is able to track every time a person is exposed to the media on a census-wide or population-wide level. As a result, the AME can reliably determine the total impression count for the media without having to extrapolate from panel data collected from a relatively limited pool of panelists within the population. Frequently, such beacon requests are implemented in connection with third-party cookies. Since the AME is a third party relative to the first party serving the media to the client device, the cookie sent to the AME in the impression request to report the occurrence of the media impression of the client device is a third-party cookie. Third-party cookie tracking is used by audience measurement servers to track access to media by client devices from first-party media servers.

Tracking impressions by tagging media with beacon instructions using third-party cookies is insufficient, by itself, to enable an AME to reliably determine the unique audience size associated with the media if the AME cannot identify the individual user associated with the third-party cookie. That is, the unique audience size cannot be determined because the collected monitoring information does not uniquely identify the person(s) exposed to the media. Under such circumstances, the AME cannot determine whether two reported impressions are associated with the same person or two separate people. The AME may set a third-party cookie on a client device reporting the monitoring information to identify when multiple impressions occur using the same device. However, cookie information does not indicate whether the same person used the client device in connection with each media impression. Furthermore, the same person may access media using multiple different devices that have different cookies so that the AME cannot directly determine when two separate impressions are associated with the same person or two different people.

Furthermore, the monitoring information reported by a client device executing the beacon instructions does not provide an indication of the demographics or other user information associated with the person(s) exposed to the associated media. To at least partially address this issue, the AME establishes a panel of users who have agreed to provide their demographic information and to have their Internet browsing activities monitored. When an individual joins the panel, that person provides corresponding detailed information concerning the person's identity and demographics (e.g., gender, race, income, home location, occupation, etc.) to the AME. The AME sets a cookie on the panelist computer that enables the AME to identify the panelist whenever the panelist accesses tagged media and, thus, sends monitoring information to the AME. Additionally or alternatively, the AME may identify the panelists using other techniques (independent of cookies) by, for example, prompting the user to login or identify themselves. While AMEs are able to obtain user-level information for impressions from panelists (e.g., identify unique individuals associated with particular media impressions), most of the client devices providing monitoring information from the tagged pages are not panelists. Thus, the identity of most people accessing media remains unknown to the AME such that it is necessary for the AME to use statistical methods to impute demographic information based on the data collected for panelists to the larger population of users providing data for the tagged media. However, panel sizes of AMEs remain small compared to the general population of users.

There are many database proprietors operating on the Internet. These database proprietors provide services to large numbers of subscribers. In exchange for the provision of services, the subscribers register with the database proprietors. Examples of such database proprietors include social network sites (e.g., Facebook, Twitter, My Space, etc.), multi-service sites (e.g., Yahoo!, Google, Axiom, Catalina, etc.), online retailer sites (e.g., Amazon.com, Buy.com, etc.), credit reporting sites (e.g., Experian), streaming media sites (e.g., YouTube, Hulu, etc.), etc. These database proprietors set cookies and/or other device/user identifiers on the client devices of their subscribers to enable the database proprietors to recognize their subscribers when their subscribers visit website(s) on the Internet domains of the database proprietors.

The protocols of the Internet make cookies inaccessible outside of the domain (e.g., Internet domain, domain name, etc.) on which they were set. Thus, a cookie set in, for example, the YouTube.com domain (e.g., a first party) is accessible to servers in theYouTube.com domain, but not to servers outside that domain. Therefore, although an AME (e.g., a third party) might find it advantageous to access the cookies set by the database proprietors, they are unable to do so. However, techniques have been developed that enable an AME to leverage media impression information collected in association with demographic information in subscriber databases of database proprietors to collect more extensive Internet usage (e.g., beyond the limited pool of individuals participating in an AME panel) by extending the impression request process to encompass partnered database proprietors and by using such partners as interim data collectors. In particular, this task is accomplished by structuring the AME to respond to impression requests from clients (who may not be a member of an audience measurement panel and, thus, may be unknown to the AME) by redirecting the clients from the AME to a database proprietor, such as a social network site partnered with the AME, using an impression response. Such a redirection initiates a communication session between the client accessing the tagged media and the database proprietor. For example, the impression response received from the AME may cause the client to send a second impression request to the database proprietor along with a cookie set by that database proprietor. In response to receiving this impression request, the database proprietor (e.g., Facebook) can access the cookie it has set on the client to thereby identify the client based on the internal records of the database proprietor.

In the event the client corresponds to a subscriber of the database proprietor (as determined from the cookie associated with the client), the database proprietor logs/records a database proprietor demographic impression in association with the client/user. As used herein, a demographic impression is an impression that can be matched to particular demographic information of a particular subscriber or registered users of the services of a database proprietor. The database proprietor has the demographic information for the particular subscriber because the subscriber would have provided such information when setting up an account to subscribe to the services of the database proprietor.

Sharing of demographic information associated with subscribers of database proprietors enables AMEs to extend or supplement their panel data with substantially reliable demographics information from external sources (e.g., database proprietors), thus extending the coverage, accuracy, and/or completeness of their demographics-based audience measurements. Such access also enables the AME to monitor persons who would not otherwise have joined an AME panel. Any web service provider having a database identifying demographics of a set of individuals may cooperate with the AME. Such web service providers may be referred to as “database proprietors” and include, for example, wireless service carriers, mobile software/service providers, social media sites (e.g., Facebook, Twitter, MySpace, etc.), online retailer sites (e.g., Amazon.com, Buy.com, etc.), multi-service sites (e.g., Yahoo!, Google, Experian, etc.), and/or any other Internet sites that collect demographic data of users and/or otherwise maintain user registration records. The use of demographic information from disparate data sources (e.g., high-quality demographic information from the panels of an audience measurement entity and/or registered user data of database proprietors) results in improved reporting effectiveness of metrics for both online and offline advertising campaigns.

The above approach to generating audience metrics by an AME depends upon the beacon requests (or tags) associated with the media to be monitored to enable an AME to obtain census wide impression counts (e.g., impressions that include the entire population exposed to the media regardless of whether the audience members are panelists of the AME). Further, the above approach also depends on third-party cookies to enable the enrichment of the census impressions with demographic information from database proprietors. However, in more recent years, there has been a movement away from the use of third-party cookies by third parties. Thus, while media providers (e.g., database proprietors) may still use first-party cookies to collect first-party data, the elimination of third-party cookies prevents the tracking of Internet media by AMEs (outside of client devices associated with panelists for which the AME has provided a meter to track Internet usage behavior). Furthermore, independent of the use of cookies, some database proprietors are moving towards the elimination of third party impression requests or tags (e.g., redirect instructions) embedded in media (e.g., beginning in 2020, third-party tags will no longer be allowed on Youtube.com and other Google Video Partner (GVP) sites). As technology moves in this direction, AMEs (e.g., third parties) will no longer be able to track census wide impressions of media in the manner they have in the past. Furthermore, AMEs will no longer be able to send a redirect request to a client accessing media to cause a second impression request to a database proprietor to associate the impression with demographic information. Thus, the only Internet media monitoring that AMEs will be able to directly perform in such a system will be with panelists that have agreed to be monitored using different techniques that do not depend on third-party cookies and/or tags.

Examples disclosed herein overcome at least some of the limitations that arise out of the elimination of third-party cookies and/or third-party tags by enabling the merging of high-quality demographic information from the panels of an AME with media impression data that continues to be collected by database proprietors. As mentioned above, while third-party cookies and/or third-party tags may be eliminated, database proprietors that provide and/or manage the delivery of media accessed online are still able to track impressions of the media (e.g., via first-party cookies and/or first-party tags). Furthermore, database proprietors are still able to associate demographic information with the impressions whenever the impressions can be matched to a particular subscriber of the database proprietor for which demographic information has been collected (e.g., when the user registered with the database proprietor). In some examples, the merging of AME panel data and database proprietor impressions data is merged in a privacy-protected cloud environment maintained by the database proprietor.

Examples disclosed herein generate computer-trained machine learning models to correct computer-generated errors in audience data, such as misattribution errors and/or non-coverage errors. Misattribution error refers to the measurement bias (e.g., generated by a computer) that occurs when a first person belonging to a first demographic group is believed to be the person associated with a media impression on a device when, in fact, a second person belonging to a second demographic group (e.g., a second demographic group different from the first demographic group) is the person for whom the media impression occurred. As used herein, non-coverage error refers to the measurement bias (e.g., generated by a computer) that occurs due to the inability of the database proprietor to recognize (e.g., identify the demographics of) a portion of the audience using network-connected devices (e.g., Internet-connected devices, mobile devices, smartphones, tablet devices, computers, Internet televisions, etc.) to view media. In examples disclosed herein, the privacy-protected cloud environment includes the capability to run computer-generated machine learning models to correct for the computer-generated errors in the audience data. However, in prior cloud environments, there is no ability to compare results between different computer-generated machine learning models to determine the best performing variant of the computer-generated machine learning models.

Examples disclosed herein generate computer-trained machine learning models in a privacy-protected cloud environment and determine performance results from the variants of the computer-trained machine learning models. Examples disclosed herein use covariates associated with user data from a database proprietor (e.g., streaming browsing category data, search browsing category data, hours of the day the user is active, etc.) to generate the computer-trained machine learning models to correct for the computer-generated errors in audience data. Examples disclosed herein generate a plurality of machine learning models with varying combinations of features and ranges for hyperparameters. Examples disclosed herein run the different combinations of computer-generated machine learning models in parallel using the user data from the database proprietor. Examples disclosed herein determine performance results for the different computer-generated machine learning models (e.g., model accuracy, demographic data accuracy, etc.) based on a comparison of the results of the computer-generated machine learning models and the audience data associated with audience measurement panelists from the AME. Examples disclosed herein aggregate the performance results for the computer-generated machine learning models and select one or more of the computer-generated machine learning models based on the performance results to use in correcting the computer-generated errors in audience data.

More particularly, FIG. 1 is a block diagram illustrating an example system 100 to enable the generation of audience measurement metrics based on the merging of data collected by a database proprietor 102 and an AME 104. More particularly, in some examples, the data includes AME panel data (that includes media impressions for panelists that are associated with high-quality demographic information collected by the AME 104) and database proprietor impressions data (which may be enriched with demographic and/or other information available to the database proprietor 102). In the illustrated example, these disparate sources of data are combined within a privacy-protected cloud environment 106 managed and/or maintained by the database proprietor 102. The privacy-protected cloud environment 106 is a cloud-based environment that enables media providers (e.g., advertisers and/or content providers) and third parties (e.g., the AME 104) to input and combine their data with data from the database proprietor 102 inside a data warehouse or data store that enables efficient big data analysis. The combining of data from different parties (e.g., different Internet domains) presents risks to the privacy of the data associated with individuals represented by the data from the different parties. Accordingly, the privacy-protected cloud environment 106 is established with privacy constraints that prevent any associated party (including the database proprietor 102) from accessing private information associated with particular individuals. Rather, any data extracted from the privacy-protected cloud environment 106 following a big data analysis and/or query is limited to aggregated information. A specific example that may be used to implement the privacy-protected cloud environment 106 is the Ads Data Hub (ADH) developed by Google LLC of Mountain View, Calif., U.S.A.

As used herein, a media impression is defined as an occurrence of access and/or exposure to media 108 (e.g., an advertisement, a movie, a movie trailer, a song, a web page banner, etc.). Examples disclosed herein may be used to monitor for media impressions of any one or more media types (e.g., video, audio, a web page, an image, text, etc.). In examples disclosed herein, the media 108 may be primary content and/or advertisements. Examples disclosed herein are not restricted for use with any particular type of media. On the contrary, examples disclosed herein may be implemented in connection with tracking impressions for media of any type or form in a network.

In the illustrated example of FIG. 1, content providers and/or advertisers distribute the media 108 via the Internet to users that access websites and/or online television services (e.g., web-based TV, Internet protocol TV (IPTV), etc.). For purposes of explanation, examples disclosed herein are described assuming the media 108 is an advertisement that may be provided in connection with particular content of primary interest to a user. In some examples, the media 108 is served by media servers managed by and/or associated with the database proprietor 102 that manages and/or maintains the privacy-protected cloud environment 106. For example, the database proprietor 102 may be Google, and the media 108 corresponds to ads served with videos accessed via Youtube.com and/or via other Google video partners (GVPs). More generally, in some examples, the database proprietor 102 includes corresponding database proprietor servers that can serve media 108 to individuals via client devices 110. In the illustrated example of FIG. 1, the client devices 110 may be stationary or portable computers, handheld computing devices, smart phones, Internet appliances, smart televisions, and/or any other type of device that may be connected to the Internet and capable of presenting media. For purposes of explanation, the client devices 110 of FIG. 1 include panelist client devices 112 and non-panelist client devices 114 to indicate that at least some individuals that access and/or are exposed to the media 108 correspond to panelists who have provided detailed demographic information to the AME 104 and have agreed to enable the AME 104 to track their exposure to the media 108. In many situations, other individuals who are not panelists will also be exposed to the media 108 (e.g., via the non-panelist client devices 114). Typically, the number of non-panelist audience members for a particular media item will be significantly greater than the number of panelist audience members. In some examples, the panelist client devices 112 may include and/or implement an audience measurement meter 115 that captures the impressions of media 108 accessed by the panelist client devices 112 (along with associated information) and reports the same to the AME 104. In some examples, the audience measurement meter 115 may be a separate device from the panelist client device 112 used to access the media 108.

In some examples, the media 108 is associated with a unique impression identifier (e.g., a consumer playback nonce (CPN)) generated by the database proprietor 102. In some examples, the impression identifier serves to uniquely identify a particular impression of the media 108. Thus, even though the same media 108 may be served multiple times, each time the media 108 is served the database proprietor 102 will generate a new and different impression identifier so that each impression of the media 108 can be distinguished from every other impression of the media. In some examples, the impression identifier is encoded into a uniform resource locator (URL) used to access the primary content (e.g., a particular YouTube video) along with which the media 108 (as an advertisement) is served. In some examples, with the impression identifier (e.g., CPN) encoded into the URL associated with the media 108, the audience measurement meter 115 extracts the identifier at the time that a media impression occurs so that the AME 104 is able to associate a captured impression with the impression identifier.

In some examples, the meter 115 may not be able to obtain the impression identifier (e.g., CPN) to associate with a particular media impression. For instance, in some examples where the panelist client device 112 is a mobile device, the meter 115 collects a mobile advertising identifier (MAID) and/or an identifier for advertisers (IDFA) that may be used to uniquely identify client devices 110 (e.g., the panelist client devices 112 being monitored by the AME 104). In some examples, the meter 115 reports the MAID and/or IDFA for the particular device associated with the meter 115 to the AME 104. The AME 104, in turn, provides the MAID and/or IDFA to the database proprietor 102 in a double blind exchange through which the database proprietor 102 provides the AME 104 with the impression identifiers (e.g., CPNs) associated with the client device 110 identified by the MAID and/or IDFA. Once the AME 104 receives the impression identifiers for the client device 110 (e.g., a particular panelist client device 112), the impression identifiers are associated with the impressions previously collected in connection with the device.

In the illustrated example, the database proprietor 102 logs each media impression occurring on any of the client devices 110 within the privacy-protected cloud environment 106. In some examples, logging an impression includes logging the time the impression occurred and the type of client device 110 (e.g., whether a desktop device, a mobile device, a tablet device, etc.) on which the impression occurred. Further, in some examples, impressions are logged along with the impression's unique impression identifier. In this example, the impressions and associated identifiers are logged in a campaign impressions database 116. The campaign impressions database 116 stores all impressions of the media 108 regardless of whether any particular impression was detected from a panelist client device 112 or a non-panelist client device 114. Furthermore, the campaign impressions database 116 stores all impressions of the media 108 regardless of whether the database proprietor 102 is able to match any particular impression to a particular subscriber of the database proprietor 102. As mentioned above, in some examples, the database proprietor 102 identifies a particular user (e.g., subscriber) associated with a particular media impression based on a cookie stored on the client device 110. In some examples, the database proprietor 102 associates a particular media impression with a user that was signed into the online services of the database proprietor 102 at the time the media impression occurred. In some examples, in addition to logging such impressions and associated identifiers in the campaign impressions database 116, the database proprietor 102 separately logs such impressions in a matchable impressions database 118. As used herein, a matchable impression is an impression that the database proprietor 102 is able to match to at least one of a particular subscriber (e.g., because the impression occurred on a client device 110 on which a user was signed into the database proprietor 102) or a particular client device 110 (e.g., based on a first-party cookie of the database proprietor 102 detected on the client device 110). In some examples, if the database proprietor 102 cannot match a particular media impression (e.g., because no user was signed in at the time the media impression occurred and there is no recognizable cookie on the associated client device 110) the impressions is omitted from the matchable impressions database 118 but is still logged in the campaign impressions database 116.

As indicated above, the matchable impressions database 118 includes media impressions (and associated unique impression identifiers) that the database proprietor 102 is able to match to a particular user that has registered with the database proprietor 102. In some examples, the matchable impressions database 118 also includes user-based covariates that correspond to the particular user to which each impression in the database was matched. As used herein, a user-based covariate refers to any item(s) of information collected and/or generated by the database proprietor 102 that can be used to identify, characterize, quantify, and/or distinguish particular users and/or their associated behavior. For example, user-based covariates may include the name, age, and/or gender of the user (and/or any other demographic information about the user) collected at the time the user registered with the database proprietor 102, and/or the relative frequency with which the user uses the different types of client device 110, the number of media items the user has accessed during a most recent period of time (e.g., the last 30 days), the search terms entered by the user during a most recent period of time (e.g., the last 30 days), feature embeddings (numerical representations) of classifications of videos viewed and/or searches entered by the user, etc. As mentioned above, the matchable database 118 also includes impressions matched to particular client devices 110 (based on first-party cookies), even when the impressions cannot be matched to particular users (based on the users being signed in at the time). In some such examples, the impressions matched to particular client devices 110 are treated as distinct users within the matchable database 118. However, as no particular user can be identified, such impressions in the matchable database 118 will not be associated with any user-based covariates.

Although only one campaign impressions database 116 is shown in the illustrated example, the privacy-protected cloud environment 106 may include any number of campaign impressions databases 116, with each database storing impressions corresponding to different media campaigns associated with one or more different advertisers (e.g., product manufacturers, service providers, retailers, advertisement servers, etc.). In other examples, a single campaign impressions database 116 may store the impressions associated with multiple different campaigns. In some such examples, the campaign impressions database 116 may store a campaign identifier in connection with each impression to identify the particular campaign to which the impression is associated. Similarly, in some examples, the privacy-protected cloud environment 106 may include one or more matchable impressions databases 118 as appropriate. Further, in some examples, the campaign impressions database 116 and the matchable impressions database 118 may be combined and/or represented in a single database.

In the illustrated example of FIG. 1, impressions occurring on the client devices 110 are shown as being reported (e.g., via network communications) directly to both the campaign impressions database 116 and the matchable impressions database 118. However, this should not be interpreted as necessarily requiring multiple separate network communications from the client devices 110 to the database proprietor 102. Rather, in some examples, notifications of impressions are collected from a single network communication from the client device 110, and the database proprietor 102 then populates both the campaign impressions database 116 and the matchable impressions database 118. In some examples, the matchable impressions database 118 is generated based on an analysis of the data in the campaign impressions database 116. Regardless of the particular process by which the two databases 116, 118 are populated with logged impressions, in some examples, the user-based covariates included in the matchable impressions database 118 may be combined with the logged impressions in the campaign impressions database 116 and stored in an enriched impressions database 120. Thus, the enriched impressions database includes all (e.g., census wide) logged impressions of the media 108 for the relevant advertising campaign and also includes all available user-based covariates associated with each of the logged impressions that the database proprietor 102 was able to match to a particular user.

As shown in the illustrated example, whereas the database proprietor 102 is able to collect impressions from both panelist client devices 112 and non-panelist client devices 114, the AME 104 is limited to collecting impressions from panelist client devices 112. In some examples, the AME 104 also collects the impression identifier associated with each collected media impression so that the collected impressions may be matched with the impressions collected by the database proprietor 102 as described further below. In the illustrated example, the impressions (and associated impression identifiers) of the panelists are stored in an AME panel data database 122 that is within an AME first party data store 124 in an AME proprietary cloud environment 126. In some examples, the AME proprietary cloud environment 126 is a cloud-based storage system (e.g., a Google Cloud Project) provided by the database proprietor 102 that includes functionality to enable interfacing with the privacy-protected cloud environment 106 also maintained by the database proprietor 102. As mentioned above, the privacy-protected cloud environment 106 is governed by privacy constraints that prevent any party (with some limited exceptions for the database proprietor 102) from accessing private information associated with particular individuals. By contrast, the AME proprietary cloud environment 126 is indicated as proprietary because it is exclusively controlled by the AME such that the AME has full control and access to the data without limitation. While some examples involve the AME proprietary cloud environment 126 being a cloud-based system that is provided by the database proprietor 102, in other examples, the AME proprietary cloud environment 126 may be provided by a third party distinct from the database proprietor 102.

While the AME 104 is limited to collected impressions (and associated identifiers) from only panelists (e.g., via the panelist client devices 112), the AME 104 is able to collect panel data that is much more robust than merely media impressions. As mentioned above, the panelist client devices 112 are associated with users that have agreed to participate on a panel of the AME 104. Participation in a panel includes the provision of detailed demographic information about the panelist and/or all members in the panelist's household. Such demographic information may include age, gender, race, ethnicity, education, employment status, income level, geographic location of residence, etc. In addition to such demographic information, which may be collected at the time a user enrolls as a panelist, the panelist may also agree to enable the AME 104 to track and/or monitor various aspects of the user's behavior. For example, the AME 104 may monitor panelists' Internet usage behavior including the frequency of Internet usage, the times of day of such usage, the websites visited, and the media exposed to (from which the media impressions are collected).

AME panel data (including media impressions and associated identifiers, demographic information, and Internet usage data) is shown in FIG. 1 as being provided directly to the AME panel data database 122 from the panelist client devices 112. However, in some examples, there may be one or more intervening operations and/or components that collect and/or process the collected data before it is stored in the AME panel data database 122. For instance, in some examples, impressions are initially collected and reported to a separate server and/or database that is distinct from the AME proprietary cloud environment 126. In some such examples, this separate server and/or database may not be a cloud-based system. Further, in some examples, such a non-cloud-based system may interface directly with the privacy-protected cloud environment 106 such that the AME proprietary cloud environment 126 may be omitted entirely.

In some examples, there may be multiple different techniques and/or methodologies used to collect the AME panel data that depends on the particular circumstances involved. For example, different monitoring techniques and/or different types of audience measurement meters 115 may be employed for media accessed via a desktop computer relative to the media accessed via a mobile computing device. In some examples, the audience measurement meter 115 may be implemented as a software application that panelists agree to install on their devices to monitor all Internet usage activity on the respective devices. In some examples, the meter 115 may prompt a user of a particular device to identify themselves so that the AME 104 can confirm the identity of the user (e.g., whether it was the mother or daughter in a panelist household). In some examples, prompting a user to self-identify may be considered overly intrusive. Accordingly, in some such examples, the circumstances surrounding the behavior of the user of a panelist client device 112 (e.g., time of day, type of content being accessed, etc.) may be analyzed to infer the identity of the user to some confidence level (e.g., the accessing of children's content in the early afternoon would indicate a relatively high probability that a child is using the device at that point in time). In some examples, the audience measurement meter 115 may be a separate hardware device that is in communication with a particular panelist client device 112 and enabled to monitor the Internet usage of the panelist client device 112.

In some examples, the processes and/or techniques used by the AME 104 to capture panel data (including media impressions and who in particular was exposed to the media) can differ depending on the nature of the panelist client device 112 through which the media was accessed. For instance, in some examples, the identity of the individual using the client device 112 may be based on the individual responding to a prompt to self-identify. In some examples, such prompts are limited to desktop client devices because such a prompt is viewed as overly intrusive on a mobile device. However, without specifically prompting a user of a mobile device to self-identify, there often is no direct way to determine whether the user is the primary user of the device (e.g., the owner of the device) or someone else (e.g., a child of the primary user). Thus, there is the possibility of misattribution of media impressions within the panel data collected using mobile devices. In some examples, to overcome the issue of misattribution in the panel data, the AME 104 may develop a machine learning model that can predict the true user of a mobile device (or any device for that matter) based on information that the AME 104 does know for certain and/or has access to. For example, inputs to the machine learning model may include the composition of the panelist household, the type (e.g., genre and/or category) of the content, the daypart or time of day when the content was accessed, etc. In some examples, the truth data used to generate and validate such a model may be collected through field surveys in which the above input features are tracked and/or monitored for a subset of panelists that have agreed to be monitored in this manner (which is more intrusive than the typical passive monitoring of content accessed via mobile devices).

As mentioned above, in some examples, the AME panel data (stored in the AME panel data database 122) is merged with the database proprietor impressions data (stored in the matchable impressions database 118) within the privacy-protected cloud environment 106 to take advantage of the combination of the disparate sets of data to generate more robust and/or reliable audience measurement metrics. In particular, the database proprietor impressions data provides the advantage of volume. That is, the database proprietor impressions data corresponds to a much larger number of impressions than the AME panel data because the database proprietor impressions data includes census wide impression information that includes all impressions collected from both the panelist client devices 112 (associated with a relatively small pool of audience members) and the non-panelist client devices 114. The AME panel data provides the advantage of high-quality demographic data for a statistically significant pool of audience members (e.g., panelists) that may be used to correct for errors and/or biases in the database proprietor impressions data.

One source of error in the database proprietor impressions data is that the demographic information for matchable users collected by the database proprietor 102 during user registration may not be truthful. In particular, in some examples, many database proprietors impose age restrictions on their user accounts (e.g., a user must be at least 13 years of age, at least 18 years of age, etc.). However, when a person registers with the database proprietor 102, the user typically self-declares their age and may, therefore, lie about their age (e.g., an 11-year-old may say they are 18 years old to bypass the age restrictions for a user account). Independent of age restrictions, a particular user may choose to enter an incorrect age for any other reason or no reason at all (e.g., a 44-year-old may choose to assert they are only 25 years old). Where a database proprietor 102 does not verify the self-declared age of users, there is a relatively high likelihood that the ages of at least some registered users of the database proprietor stored in the matchable impressions database 118 (as a particular user-based covariate) are inaccurate. Further, it is possible that other self-declared demographic information (e.g., gender, race, ethnicity, income level, etc.) may also be falsified by users during registration. In some examples, demographic information for some registered users may be missing (e.g., registered users elect to not submit/declare certain demographic information). Mis-represented and/or missing demographic information from subscriber accounts of registered users of the database proprietor 102 results in inaccurate demographic-based audience measurements for accessed media. As described further below, the AME panel data (which contains reliable demographic information about the panelists) can be used to correct for inaccurate demographic information in the database proprietor impressions data.

Another source of error in the database proprietor impressions data is based on the concept of misattribution, which arises in situations where multiple different people use the same client device 110 to access media. In some examples, the database proprietor 102 associates a particular impression to a particular user based on the user being signed into a platform provided by the database proprietor. For example, if a particular person signs into their Google account and begins watching a YouTube video on a particular client device 110, that person will be attributed with an impression for an ad served during the video because the person was signed in at the time. However, there may be instances where the person finishes using the client device 110 but does not sign out of his or her Google account. Thereafter, a second different person (e.g., a different member in the family of the first person) begins using the client device 110 to view another YouTube video. Although the second person is now accessing media via the client device 110, ad impressions during this time will still be attributed to the first person because the first person is the one who is still indicated as being signed in. Thus, there is likely to be circumstances where the actual person exposed to media 108 is misattributed to a different registered user of the database proprietor 102. The AME panel data (which includes an indication of the actual person using the panelist client devices 112 at any given moment) can be used to correct for misattribution in the demographic information in the database proprietor impressions data. As mentioned above, in some situations, the AME panel data may itself include misattribution errors. Accordingly, in some examples, the AME panel data may first be corrected for misattribution before the AME panel data is used to correct misattribution in the database proprietor impressions data. An example methodology to correct for misattribution in the database proprietor impressions data is described in Singh et al., U.S. Pat. No. 10,469,903, which is hereby incorporated herein by reference in its entirety.

Additionally, examples disclosed herein use covariates associated with user data from a database proprietor (e.g., streaming browsing category data, search browsing category data, hours of the day the user is active, etc.) to generate the computer-trained machine learning models to correct for misattribution in the database proprietor impressions data. Examples disclosed herein generate a plurality of machine learning models with varying combinations of features and ranges for hyperparameters and run the different combinations of computer-generated machine learning models in parallel using the user data from the database proprietor. Example disclosed herein determine performance results for the different computer-generated machine learning models and select one or more of the computer-generated machine learning models based on the performance results to use in correcting the computer-generated misattribution.

Another problem with the database proprietor impressions data is that of non-coverage. Non-coverage refers to impressions recorded by the database proprietor 102 that cannot be matched to a particular registered user of the database proprietor 102. The inability of the database proprietor 102 to match a particular impression to a particular user can occur for several reasons including that the user is not signed in at the time of the media impression, that the user has not established an account with the database proprietor 102, that the user has enabled Limited Ad Tracking (LAT) to prevent the user account from being associated with ad impressions, or that the content associated with the media being monitored corresponds to children's content (for which user-based tracking is not performed). While the inability of the database proprietor 102 to match and assign a particular impression to a particular user is not necessarily an error in the database proprietor impressions data, it does undermine the ability to reliably estimate the total unique audience size for (e.g., the number of unique individuals that were exposed to) a particular media item. For example, assume that the database proprietor 102 records a total of 11,000 impressions for media 108 in a particular advertising campaign. Further assume that of those 11,000 impressions, the database proprietor 102 is able to match 10,000 impressions to a total of 5,000 different users (e.g., each user was exposed to the media on average 2 times) but is unable to match the remaining 1,000 impressions to particular users. Relying solely on the database proprietor impressions data, in this example, there is no way to determine whether the remaining 1,000 impressions should also be attributed to the 5,000 users already exposed at least once to the media 108 (for a total audience size of 5,000 people) or if one or more of the remaining 1,000 impressions should be attributed to other users not among the 5,000 already identified (for a total audience size of up to 6,000 people (if every one of the 1,000 impressions was associated with a different person not included in the matched 5,000 users)). In some examples disclosed herein, the AME panel data can be used to estimate the distribution of impressions across different users associated with the non-coverage portion of impressions in the database proprietor impressions data to thereby estimate a total audience size for the relevant media 108. In some examples disclosed herein, a plurality of computer-generated machine learning models with different combinations of features and hyperparameters are run using the user data from the database proprietor. Examples disclosed herein determine performance results for the different computer-generated machine learning models based on a comparison of the results of the computer-generated machine learning models and the audience data associated audience measurement panelists from the AME. Examples disclosed herein select one or more of the computer-generated machine learning models based on the performance results to use in correcting the computer-generated errors in audience data.

Another confounding factor to the estimation of the total unique audience size for media based on the database proprietor impressions data is the existence of multiple user accounts of a single user. More particular, in some situations a particular individual may establish multiple accounts with the database proprietor 102 for different purposes (e.g., a personal account, a work account, a joint account shared with other individuals, etc.). Such a situation can result in a larger number of different users being identified as audience members to media 108 than the actual number of individuals exposed to the media 108. For example, assume that a particular person registers three user accounts with the database proprietor 102 and is exposed to the media 108 once while signed into each of the three different accounts for a total of three impressions. In this scenario, the database proprietor 102 would match each impression to a different user based on the different user accounts making it appear that three different people were exposed to the media 108 when, in fact, only one person was exposed to the media three different times. Examples disclosed herein use the AME panel data in conjunction with the database proprietor impressions data to estimate an actual unique audience size from the potentially inflated number of apparently unique users exposed to the media 108.

In the illustrated example of FIG. 1, the AME panel data is merged with the database proprietor impressions data by an example data matching analyzer 128. In some examples, the data matching analyzer 128 implements an application programming interface (API) that takes the disparate datasets and matches users in the database proprietor impressions data with panelists in the AME panel data. In some examples, users are matched with panelists based on the unique impression identifiers (e.g., CPNs) collected in connection with the media impressions logged by both the database proprietor 102 and the AME 104. The combined data is stored in an intermediary merged data database 130 within an AME privacy-protected data store 132. The data in the intermediary merged data database 130 is referred to as “intermediary” because it is at an intermediate stage in the processing because it includes AME panel data that has been enhanced and/or combined with the database proprietor impressions data but has not yet be corrected or adjusted to account for the sources of error and/or bias in the database proprietor impressions data as outlined above.

In some examples, the AME intermediary merged data is analyzed by an adjustment factor analyzer 134 to calculate adjustment or calibration factors that may be stored in an adjustment factors database 136 within an AME output data store 138 of the AME proprietary cloud environment 126. In some examples, the adjustment factor analyzer 134 calculates different types of adjustment factors to account for different types of errors and/or biases in the database proprietor impressions data. For instance, a multi-account adjustment factor corrects for the situation of a single user accessing media using multiple different user accounts associated with the database proprietor 102. A signed-out adjustment factor corrects for non-coverage associated with users that access media while signed out of their account associated with the database proprietor 102 (so that the database proprietor 102 is unable to associate the impression with the users). In some examples, the adjustment factor analyzer 134 is able to directly calculate the multi-account adjustment factor and the signed-out adjustment factor in a deterministic manner.

While the multi-account adjustment factors and the signed-out adjustment factors may be deterministically calculated, correcting for falsified or otherwise incorrect demographic information (e.g., incorrectly self-declared ages) of registered users of the database proprietor 102 cannot be solved in such a direct and deterministic manner. Rather, in some examples, computer-generated machine learning models are developed to analyze and predict the correct demographics (e.g., ages) of registered users of the database proprietor 102. Specifically, as shown in FIG. 1, the privacy-protected cloud environment 106 implements a model generator 140 to generate computer-generated machine learning models using the AME intermediary merged data (stored in the AME intermediary merged data database 130) as inputs. More particularly, in some examples, self-declared demographics (e.g., the self-declared age) of users of the database proprietor 102, along with other covariates associated with the users, are used as the input variables or features used to train the computer-generated machine learning models to predict the correct demographics (e.g., correct age) of the users as validated by the AME panel data, which serves as the truth data or training labels for the model generation. The example model generator 140 determines performance results for the different computer-generated machine learning models (e.g., model accuracy, demographic data accuracy, etc.) based on a comparison of the results of the computer-generated machine learning models and the audience data associated with audience measurement panelists from the AME. After the different computer-generated machine learning models have been trained and validated based on the AME panel data, the computer-generated machine learning models and associated performance results are stored in an example demographic correction models database 142. In some examples, different demographic correction model(s) may be developed to correct for different types of demographic information that needs correcting. For instance, in some examples, a first model can be used to correct the self-declared age of users of the database proprietor 102 and a second model can be used to correct the self-declared gender of the users.

As mentioned above, there are many different types of covariates collected and/or generated by the database proprietor 102. In some examples, the covariates provided by the database proprietor 102 may include a certain number (e.g., 100) of the top search result click entities and/or video watch entities for every user during a most recent period of time (e.g., for the last month). These entities are integer identifiers (IDs) that map to a knowledge graph of all entities for the search result clicks and/or videos watched. That is, as used in this context, an entity corresponds to a particular node in a knowledge graph maintained by the database proprietor 102. In some examples, the total number of unique IDs in the knowledge graph may number in the tens of millions. More particularly, for example, YouTube videos are classified across roughly 20 million unique video entity IDs and Google search results are classified across roughly 25 million unique search result entity IDs. In addition to the top search result click entities and/or video watch entities, the database proprietor 102 may also provide embeddings for these entities. An embedding is a numerical representation (e.g., a vector array of values) of some class of similar objects, images, words, and the like. For example, a particular user that frequently searches for and/or views cat videos may be associated with a feature embedding representative of the class corresponding to cats. Thus, feature embeddings translate relatively high dimensional vectors of information (e.g., text strings, images, videos, etc.) into a lower dimensional space to enable the classification of different but similar objects.

In some examples, multiple embeddings may be associated with each search result click entity and/or video watch entity. Accordingly, assuming the top 100 search result entities and video watch entities are provided among the covariates and that 16-dimension embeddings are provided for each such entity, this results in a 100×16 matrix of values for every user, which may be too much data to process during generation of the demographic correction models as described above. Accordingly, in some examples, the dimensionality of the matrix is reduced to a more manageable size to be used as an input feature for the demographic correction model generation.

In some examples, a process is implemented to track different computer-generated machine learning model experiments over time to achieve high quality (e.g., accurate) models and also for auditing purposes. Accomplishing this objective within the context of the privacy-protected cloud environment 106 presents several unique challenges because the model features (e.g., inputs and hyperparameters) and model performance (e.g., accuracy) are stored separately to satisfy the privacy constraints of the environment.

In some examples, a model analyzer 144 may implement and/or use one or more computer-generated machine learning models to generate predictions and/or inferences as to the actual demographics (e.g., actual ages) of users associated with media impressions logged by the database proprietor 102. In examples disclosed herein, the model analyzer 144 obtains the computer-generated machine learning models and associated performance results stored in the demographic correction models database 142. The example model analyzer 144 aggregates the performance results for the computer-generated machine learning models and selects one or more of the computer-generated machine learning models based on the performance results to use in correcting human errors (e.g., errors from users self-declaring/entering inaccurate demographic information such as, age, gender, etc.) and/or computer-generated errors (e.g., misattribution error(s), non-coverage error(s), etc.) in the demographic information for users associated with the impressions from the database proprietor 102. In some examples, as shown in FIG. 1, the model analyzer 144 uses the selected computer-generated machine learning model(s) from the demographic correction models database 142 to analyze the impressions in the enriched impressions database 120 that were matched to a particular user of the database proprietor 102. The inferred demographic (e.g., age) for each user may be stored in a model inferences database 146 for subsequent use, retrieval, and/or analysis. Additionally or alternatively, in some examples, the model analyzer 144 uses the selected computer-generated machine learning model(s) from the demographic correction models database 142 to analyze the entire user base of the database proprietor 102 regardless of whether the users are matched to any particular media impressions. After inferring the correct demographic (e.g., age) for each user, the inferences are stored in the model inferences database 146. In some such examples, when the users matched to particular impressions are to be analyzed (e.g., the users matched to impressions in the enriched impressions database 120), the model analyzer 144 merely extracts the inferred demographic assignment to each relevant user in the enriched impressions database 120 that matches with one or more media impressions.

As described above, in some examples, the database proprietor 102 may identify a particular user as corresponding to a particular impression based on the user being signed into the database proprietor 102. However, there are circumstances where the individual corresponding to the user account is not the actual person that was exposed to the relevant media. Accordingly, merely inferring a correct demographic (e.g., age) of the user associated with the signed in user account may not be the correct demographic of the actual person to which a particular media impression should be attributed. In other words, whereas the AME panelist data and the database proprietor impressions data is matched at the impression level, demographic correction is implemented at the user level. Therefore, before generating the demographic correction model, a method to reduce logged impressions to individual users is first implemented so that the demographic correction model can be reliably implemented.

With inferences made to correct for inaccurate demographic information of database proprietor users (e.g., falsified self-declared ages) and stored in the model inferences database 146, the AME 104 may be interested in extracting audience measurement metrics based on the corrected data. However, as mentioned above, the data contained inside the privacy-protected cloud environment 106 is subject to privacy constraints. In some examples, the privacy constraints ensure that the data can only be extracted for review and/or analysis in aggregate so as to protect the privacy of any particular individual represented in the data (e.g., a panelist of the AME 104 and/or a registered user of the database proprietor 102). Accordingly, in some examples, a data aggregator 148 aggregates the audience measurement data associated with particular media campaigns before the data is provided to an aggregated campaign data database 150 in the AME output data store 138 of the AME proprietary cloud environment 126.

The data aggregator 148 may aggregate data in different ways for different types of audience measurement metrics. For instance, at the highest level, the aggregated data may provide the total impression count and total number of users (e.g., estimated audience size) exposed to the media 108 for a particular media campaign. As mentioned above, the total number of users reported by the data aggregator 148 is based on the total number of unique user accounts matched to impressions but does not include the individuals associated with impressions that were not matched to a particular user (e.g., non-coverage). However, the total number of unique user accounts does not account for the fact that a single individual may correspond to more than one user account (e.g., multi-account users), and does not account for situations where a person other than a signed-in user was exposed to the media 108 (e.g., misattribution). These errors in the aggregated data may be corrected based on the adjustment factors stored in the adjustment factors database 136. Further, in some examples, the aggregated data may include an indication of the demographic composition of the users represented in the aggregated data (e.g., number of males vs females, number of users in different age brackets, etc.).

Additionally or alternatively, in some examples, the data aggregator 148 may provide aggregated data that is associated with a particular aspect of a media campaign. For instance, the data may be aggregated based on particular sites (e.g., all media impressions served on YouTube.com). In other examples, the data may be aggregated based on placement information (e.g., aggregated based on particular primary content videos accessed by users when the media advertisement was served). In other examples, the data may be aggregated based on device type (e.g., impressions served via a desktop computer versus impressions served via a mobile device). In other examples, the data may be aggregated based on a combination of one or more of the above factors and/or based on any other relevant factor(s).

In some examples, the privacy constraints imposed on the data within the privacy-protected cloud environment 106 include a limitation that data cannot be extracted (even when aggregated) for less than a threshold number of individuals (e.g., 50 individuals). Accordingly, if the particular metric being sought includes less than the threshold number of individuals, the data aggregator 148 will not provide such data. For instance, if the threshold number of individuals is 50 but there are only 46 females in the age range of 18-25 that were exposed to particular media 108, the data aggregator 148 would not provide the aggregate data for females in the 18-25 age bracket. Such privacy constraints can leave gaps in the audience measurement metrics, particularly in locations where the number of panelists is relatively small. Accordingly, in some examples, when audience measurement is not available for a particular demographic segment of interest in a particular region (e.g., a particular country), the audience measurement metrics in one or more comparable region(s) may be used to impute the metrics for the missing data in the first region of interest. In some examples, the particular metrics imputed from comparable regions is based on a comparison of audience metrics for which data is available in both regions. For instance, while data for females in the 18-25 bracket may be unavailable, assume that data for females in the 26-35 age bracket is available. The metrics associated with the 26-35 age bracket in the region of interests may be compared with metrics for the 26-35 age bracket in other regions and the regions with the closest metrics to the region of interest may be selected for use in calculating imputation factor(s).

As shown in the illustrated example, both the adjustment factors database 136 and the aggregated campaigns data database 150 are included within the AME output data store 138 of the AME proprietary cloud environment 126. As mentioned above, in some examples, the AME proprietary cloud environment 126 is provided by the database proprietor 102 and enables data to be provided to and retrieved from the privacy-protected cloud environment. In some examples, the aggregated campaign data and the adjustment factors are subsequently transferred to a separate computing apparatus 152 of the AME 104 for analysis by an audience metrics analyzer 154. In some examples, the separate computing apparatus may be omitted with its functionality provided by the AME proprietary cloud environment 126. In other examples, the AME proprietary cloud environment 126 may be omitted with the adjustment factors and the aggregated data provided directly to the computing apparatus 152. Further, in this example, the AME panel data database 122 is within the AME first party data store 124, which is shown as being separate from the AME output data store 138. However, in other examples, the AME first party data store 124 and the AME output data store 138 may be combined.

In the illustrated example of FIG. 1, the audience metrics analyzer 154 applies the adjustment factors to the aggregated data to correct for errors in the data including misattribution, non-coverage, and multi-account users. The output of the audience metrics analyzer 154 corresponds to the final calibrated data of the AME 104 and is stored in a final calibrated data database 156. In this example, the computing apparatus 152 also includes a report generator 158 to generate reports based on the final calibrated data.

FIG. 2 is a block diagram of the example model generator 140 of FIG. 1. The example model generator 140 of FIG. 2 includes an example feature interface 202, an example hyperparameter interface 204, an example query selector 206, and an example query generator 208.

In the illustrated example of FIG. 2, the example feature interface 202 accesses candidate features (e.g., from memory, a receive buffer, etc.) to use as inputs for generating the computer-generated machine learning models. In some examples, the feature interface 202 accesses covariates associated with user data from the database proprietor 102. In examples disclosed herein, covariates are used as candidate features for generating the computer-generated machine learning models. In some examples, the covariates represent information collected and/or generated by the database proprietor 102 that are used to identify, characterize, quantify, and/or distinguish particular users and/or their associated behavior. For example, the covariates may include the name, age, and/or gender of the user (and/or any other demographic information about the user) collected at the time the user registered with the database proprietor 102, and/or the relative frequency with which the user uses the different types of client device 110, the number of media items the user has accessed during a most recent period of time (e.g., the last 30 days), the search terms entered by the user during a most recent period of time (e.g., the last 30 days), feature embeddings (numerical representations) of classifications of videos viewed and/or searches entered by the user, etc. In some examples, the feature interface 202 includes a range of candidate features (e.g., 20 to 100 candidate features) for the computer-generated machine learning models based on the covariates.

In the illustrated example, the example hyperparameter interface 204 accesses candidate hyperparameters (e.g., from memory, a receive buffer, etc.) to use as inputs for generating the computer-generated machine learning models. In some examples, the hyperparameter interface 204 accesses the different candidate hyperparameters available to the computer-generated machine learning models from the database proprietor 102. In some examples, the candidate hyperparameters available from the database proprietor 102 can include topology of a neural network, size of a neural network, learning rate of the neural network, batch size of the neural network, etc. In some examples, the hyperparameter interface 204 accesses eight different hyperparameters with adjustable values from the database proprietor 102 to input in the computer-generated machine learning models.

In the illustrated example of FIG. 2, the example query selector 206 selects a plurality of features from the example feature interface 202 and selects a range of hyperparameters from the example hyperparameter interface 204. The example query selector 206 selects features from the feature interface 202 on which to run different combinations of machine learning models. In some examples, the query selector 206 selects eight different features from the candidate features via the feature interface 202 for each of the computer-generated machine learning models. For example, the query selector 206 selects a first set of eight features from the candidate features for a first computer-generated machine learning model and selects a second set of eight features from the candidate features for a second computer-generated machine learning model, where the first set of eight features and the second set of eight features may differ by any number of features (e.g., one different feature, eight different features, etc.). However, the example query selector 206 can select any number of different features for the computer-generated machine learning models. The example query selector 206 selects ranges for the hyperparameters obtained via the hyperparameter interface 204 to provide to the computer-generated machine learning models. The example query selector 206 selects different combinations of the features via the feature interface 202 and ranges of the hyperparameters via the hyperparameter interface 204 for each of the different machine learning models. In some examples, the query selector 206 selects a plurality of combinations of features and hyperparameters to be used to correct computer-generated data errors as disclosed herein. In some examples, the plurality of combinations of features and hyperparameters include every possible combination of features and hyperparameters available via the example feature interface 202 and the example hyperparameter interface 204. In such examples, the different combinations of candidate features and hyperparameters are evaluated using the computer-generated machine learning models so that all of the generated results of those models can be evaluated relative to one another. In this manner, one or more of the combinations of candidate features and hyperparameters can be selected based on such performance evaluations (e.g., select one or more of the combinations of candidate features and hyperparameters that achieve relatively better performance than other ones of the combinations of candidate features and hyperparameters) for use in correcting any human-generated errors (e.g., errors from users self-declaring inaccurate demographic information such as, age, gender, etc.) and/or any computer-generated errors in the user demographic information (e.g., misattribution, non-coverage, etc.).

The example query generator 208 generates a plurality of different computer-generated machine learning models based on the plurality of selected combinations of the sets of features and ranges of hyperparameters from the example query selector 206. The example query generator 208 initiates the training of the plurality of computer-generated machine learning models based on demographic data of audience measurement panelists that are also subscribers of the database proprietor 102. In the illustrated example, the demographic data is obtained from user accounts of the database proprietor 102 in the privacy-protected cloud environment 106. In some examples, the query generator 208 uses the demographic data of the users from the database proprietor 102, the selected features, and the selected ranges of hyperparameters to train the plurality of computer-generated machine learning models. In some examples, the query generator 208 triggers the training of the computer-generated machine learning models in parallel. In some examples, manually running and training only one of the computer-generated machine learning models could take up to ten minutes. However, in the illustrated example, the example query generator 208 triggers the running and training of all of the computer-generated machine learning models in parallel to allow for running and training all of the computer-generated machine learning models in about ten minutes. The example query generator 208 generates the demographic results of each of the computer-generated machine learning models based on the training and running of the computer-generated machine learning models.

In the illustrated example of FIG. 2, the example analytics controller 210 generates performance results for all of the computer-generated machine learning models from the example query generator 208. The example analytics controller 210 compares the results (e.g., demographic information for users) from the training of the plurality of computer-generated machine learning models to the demographic data from audience measurement panelists from the AME panel data. In some examples, the demographic data from audience measurement panelists from the AME panel data is used to validate the demographic results of the computer-generated machine learning models (e.g., the AME panel data serves as the truth data). In some examples, the analytics controller 210 obtains demographic data of the audience measurement panelist who access media via panelist client devices that correspond to the subscribers of the database proprietor 102 to determine the performance of each of the computer-generated machine learning models. In some examples, the performance results include model accuracy, demographic accuracy, etc. In some examples, the example analytics controller 210 stores the computer-generated machine learning models and the corresponding performance results in the example demographic correction models database 142 of FIG. 1.

FIG. 3 is a block diagram illustrating the example model analyzer 144 of FIG. 1. The example model analyzer 144 of FIG. 3 includes an example query results interface 302, an example data aggregation controller 304, and an example model selector 306.

The example query results interface 302 obtains the computer-generated machine learning models and corresponding performance results from the example analytics controller 210 of FIG. 2. In some examples, the example query results interface 302 obtains the computer-generated machine learning models and corresponding performance results stored in the example demographic correction models database 142 of FIG. 1.

In the illustrated example of FIG. 3, the example data aggregation controller 304 runs a query to merge the individual performance results for the computer-generated machine learning models to aggregate the performance results. The example data aggregation controller 304 generates aggregate results of the performance results from the plurality of computer-generated machine learning models. In some examples, the data aggregation controller 304 aggregates the performance results into a tabular format that is easier to read and interpret when analyzing the performance results of the computer-generated machine learning models.

In the illustrated example of FIG. 3, the example model selector 306 compares the performance results of the computer-generated machine learning models based on the aggregated performance results from the example data aggregation controller 304. The example model selector 306 selects one of the computer-generated machine learning models based on the performance results. In some examples, the example model selector 306 compares the performance results across the computer-generated machine learning models to select a computer-generated machine learning model with the relatively best performance. In some examples, the example model selector 306 determines a computer-generated machine learning model has relatively the best performance when the aggregated performance results for the computer-generated machine learning model are higher in value (e.g., a higher value for model accuracy, a higher value for demographic accuracy, etc.) compared to the aggregated performance results for the remaining computer-generated machine learning models. In some examples, the example model selector 306 determines a computer-generated machine learning model has relatively the best performance based on a combination of all of the aggregated performance results for the computer-generated machine learning model. In some examples, the different performance results for the computer-generated machine learning model can be weighted together to determine a computer-generated machine learning model with relatively the best performance. In some examples, the example model selector 306 determines a computer-generated machine learning model with the relatively best performance based on a combination of the performance metrics determined by the example analytics controller 210 of FIG. 2 (e.g., model accuracy, demographic accuracy, etc.). The example model selector 306 determines which combination of features and hyperparameters yielded the relatively best performance results among the computer-generated machine learning models. The example model selector 306 enables the comparison of the performance results from all of the computer-generated machine learning models in a fraction of the time relative to comparing the performance results manually.

In some examples, the example model selector 306 applies the selected computer-generated machine learning model to the impressions from the database proprietor 102 to correct for any human errors (e.g., errors from users self-declaring inaccurate demographic information such as, age, gender, etc.) and/or any computer-generated errors in the user demographic information (e.g., misattribution, non-coverage, etc.). In some examples, the example model selector 306 uses the selected computer-generated machine learning model to analyze the impressions in the example enriched impressions database 120 of FIG. 1 that were matched to a particular user of the database proprietor 102. The demographic information (e.g., age, gender, etc.) from applying the selected computer-generated machine learning for each user may be stored in the example model inferences database 146 of FIG. 1 for subsequent use, retrieval, and/or analysis. In some examples, the example model selector 306 uses the selected computer-generated machine learning model to analyze the entire user base of the database proprietor 102 regardless of whether the users are matched to any particular media impressions. After inferring the correct demographic information (e.g., age, gender, etc.) for each user using the selected computer-generated machine learning model, the example model selector 306 stores the inferences in the example model inferences database 146 of FIG. 1. In some such examples, when the users matched to particular impressions are to be analyzed (e.g., the users matched to impressions in the enriched impressions database 120), the example model selector 306 extracts the inferred demographic assignment to each relevant user in the enriched impressions database 120 that matches with one or more media impressions. In some examples, the model selector 306 can infer other information in addition to or instead of the demographic information for each user using the selected computer-generated machine learning model to correct for any other human errors and/or any computer-generated errors in the media impressions.

While example manners of implementing the model generator 140 and the model analyzer 144 of FIG. 1 are illustrated in FIGS. 2 and 3, one or more of the elements, processes and/or devices illustrated in FIGS. 2 and 3 may be combined, divided, re-arranged, omitted, eliminated and/or implemented in any other way. Further, the example feature interface 202, the example hyperparameter interface 204, the example query selector 206, the example query generator 208, the example analytics controller 210, the example query results interface 302, the example data aggregation controller 304, the example model selector 306 and/or, more generally, the example model generator 140 and the example model analyzer 144 of FIGS. 2 and 3 may be implemented by hardware, software, firmware and/or any combination of hardware, software and/or firmware. Thus, for example, any of the example feature interface 202, the example hyperparameter interface 204, the example query selector 206, the example query generator 208, the example analytics controller 210, the example query results interface 302, the example data aggregation controller 304, the example model selector 306 and/or, more generally, the example model generator 140 and the example model analyzer 144 could be implemented by one or more analog or digital circuit(s), logic circuits, programmable processor(s), programmable controller(s), graphics processing unit(s) (GPU(s)), digital signal processor(s) (DSP(s)), application specific integrated circuit(s) (ASIC(s)), programmable logic device(s) (PLD(s)) and/or field programmable logic device(s) (FPLD(s)). When reading any of the apparatus or system claims of this patent to cover a purely software and/or firmware implementation, at least one of the example feature interface 202, the example hyperparameter interface 204, the example query selector 206, the example query generator 208, the example analytics controller 210, the example query results interface 302, the example data aggregation controller 304, and/or the example model selector 306 is/are hereby expressly defined to include a non-transitory computer readable storage device or storage disk such as a memory, a digital versatile disk (DVD), a compact disk (CD), a Blu-ray disk, etc. including the software and/or firmware. Further still, the example model generator 140 and the example model analyzer 144 of FIGS. 2 and 3 may include one or more elements, processes and/or devices in addition to, or instead of, those illustrated in FIGS. 2 and 3, and/or may include more than one of any or all of the illustrated elements, processes and devices. As used herein, the phrase “in communication,” including variations thereof, encompasses direct communication and/or indirect communication through one or more intermediary components, and does not require direct physical (e.g., wired) communication and/or constant communication, but rather additionally includes selective communication at periodic intervals, scheduled intervals, aperiodic intervals, and/or one-time events.

Flowcharts representative of example hardware logic, machine readable instructions, hardware implemented state machines, and/or any combination thereof for implementing the model generator 140 and the model analyzer 144 of FIGS. 2 and 3 are shown in FIGS. 4 and 5. The machine readable instructions may be one or more executable programs or portion(s) of an executable program for execution by a computer processor and/or processor circuitry, such as the processor 612 shown in the example processor platform 600 discussed below in connection with FIG. 6. The program may be embodied in software stored on a non-transitory computer readable storage medium such as a CD-ROM, a floppy disk, a hard drive, a DVD, a Blu-ray disk, or a memory associated with the processor 612, but the entire program and/or parts thereof could alternatively be executed by a device other than the processor 612 and/or embodied in firmware or dedicated hardware. Further, although the example program is described with reference to the flowcharts illustrated in FIGS. 4 and 5, many other methods of implementing the example model generator 140 and the example model analyzer 144 may alternatively be used. For example, the order of execution of the blocks may be changed, and/or some of the blocks described may be changed, eliminated, or combined. Additionally or alternatively, any or all of the blocks may be implemented by one or more hardware circuits (e.g., discrete and/or integrated analog and/or digital circuitry, an FPGA, an ASIC, a comparator, an operational-amplifier (op-amp), a logic circuit, etc.) structured to perform the corresponding operation without executing software or firmware. The processor circuitry may be distributed in different network locations and/or local to one or more devices (e.g., a multi-core processor in a single machine, multiple processors distributed across a server rack, etc.).

The machine readable instructions described herein may be stored in one or more of a compressed format, an encrypted format, a fragmented format, a compiled format, an executable format, a packaged format, etc. Machine readable instructions as described herein may be stored as data or a data structure (e.g., portions of instructions, code, representations of code, etc.) that may be utilized to create, manufacture, and/or produce machine executable instructions. For example, the machine readable instructions may be fragmented and stored on one or more storage devices and/or computing devices (e.g., servers) located at the same or different locations of a network or collection of networks (e.g., in the cloud, in edge devices, etc.). The machine readable instructions may require one or more of installation, modification, adaptation, updating, combining, supplementing, configuring, decryption, decompression, unpacking, distribution, reassignment, compilation, etc. in order to make them directly readable, interpretable, and/or executable by a computing device and/or other machine. For example, the machine readable instructions may be stored in multiple parts, which are individually compressed, encrypted, and stored on separate computing devices, wherein the parts when decrypted, decompressed, and combined form a set of executable instructions that implement one or more functions that may together form a program such as that described herein.

In another example, the machine readable instructions may be stored in a state in which they may be read by processor circuitry, but require addition of a library (e.g., a dynamic link library (DLL)), a software development kit (SDK), an application programming interface (API), etc. in order to execute the instructions on a particular computing device or other device. In another example, the machine readable instructions may need to be configured (e.g., settings stored, data input, network addresses recorded, etc.) before the machine readable instructions and/or the corresponding program(s) can be executed in whole or in part. Thus, machine readable media, as used herein, may include machine readable instructions and/or program(s) regardless of the particular format or state of the machine readable instructions and/or program(s) when stored or otherwise at rest or in transit.

The machine readable instructions described herein can be represented by any past, present, or future instruction language, scripting language, programming language, etc. For example, the machine readable instructions may be represented using any of the following languages: C, C++, Java, C#, Perl, Python, JavaScript, HyperText Markup Language (HTML), Structured Query Language (SQL), Swift, etc.

As mentioned above, the example processes of FIGS. 4 and 5 may be implemented using executable instructions (e.g., computer and/or machine readable instructions) stored on a non-transitory computer and/or machine readable medium such as a hard disk drive, a flash memory, a read-only memory, a compact disk, a digital versatile disk, a cache, a random-access memory and/or any other storage device or storage disk in which information is stored for any duration (e.g., for extended time periods, permanently, for brief instances, for temporarily buffering, and/or for caching of the information). As used herein, the term non-transitory computer readable medium is expressly defined to include any type of computer readable storage device and/or storage disk and to exclude propagating signals and to exclude transmission media.

“Including” and “comprising” (and all forms and tenses thereof) are used herein to be open ended terms. Thus, whenever a claim employs any form of “include” or “comprise” (e.g., comprises, includes, comprising, including, having, etc.) as a preamble or within a claim recitation of any kind, it is to be understood that additional elements, terms, etc. may be present without falling outside the scope of the corresponding claim or recitation. As used herein, when the phrase “at least” is used as the transition term in, for example, a preamble of a claim, it is open-ended in the same manner as the term “comprising” and “including” are open ended. The term “and/or” when used, for example, in a form such as A, B, and/or C refers to any combination or subset of A, B, C such as (1) A alone, (2) B alone, (3) C alone, (4) A with B, (5) A with C, (6) B with C, and (7) A with B and with C. As used herein in the context of describing structures, components, items, objects and/or things, the phrase “at least one of A and B” is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and (3) at least one A and at least one B. Similarly, as used herein in the context of describing structures, components, items, objects and/or things, the phrase “at least one of A or B” is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and (3) at least one A and at least one B. As used herein in the context of describing the performance or execution of processes, instructions, actions, activities and/or steps, the phrase “at least one of A and B” is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and (3) at least one A and at least one B. Similarly, as used herein in the context of describing the performance or execution of processes, instructions, actions, activities and/or steps, the phrase “at least one of A or B” is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and (3) at least one A and at least one B.

As used herein, singular references (e.g., “a”, “an”, “first”, “second”, etc.) do not exclude a plurality. The term “a” or “an” item, as used herein, refers to one or more of that item. The terms “a” (or “an”), “one or more”, and “at least one” can be used interchangeably herein. Furthermore, although individually listed, a plurality of means, elements or method actions may be implemented by, e.g., a single unit or processor. Additionally, although individual features may be included in different examples or claims, these may possibly be combined, and the inclusion in different examples or claims does not imply that a combination of features is not feasible and/or advantageous.

FIG. 4 is a flowchart representative of machine readable instructions 400 which may be executed to implement the example model generator 140 of FIGS. 1 and/or 2. The example instructions 400 begin at block 402 at which the example query selector 206 (FIG. 2) selects features on which to run different combination of models. In some examples, the example query selector 206 selects a plurality of features via the example feature interface 202 of FIG. 2. In some examples, the query selector 206 selects eight different features from the candidate features via the feature interface 202 for each of the computer-generated machine learning models. For example, the query selector 206 selects a first set of eight features from the candidate features for a first computer-generated machine learning model and selects a second set of eight features from the candidate features for a second computer-generated machine learning, where one or more of the first set of eight features may differ in type from one or more of the second set of eight features (e.g., one different feature type, eight different feature types, etc.). However, the example query selector 206 can select any number of different features for the computer-generated machine learning models.

At block 404, the example query selector 206 selects a range for the hyperparameters to provide the models. In some examples, the example query selector 206 selects a range of hyperparameters via the example hyperparameter interface 204. The example query selector 206 selects ranges for the hyperparameters via the hyperparameter interface 204 to provide the computer-generated machine learning models. The example query selector 206 selects different combinations of the features from the feature interface 202 and ranges of the hyperparameters from the hyperparameter interface 204 for each of the different machine learning models. In some examples, the query selector 206 selects every possible combination of features and hyperparameters available from the example feature interface 202 and the example hyperparameter interface 204.

At block 406, the example query generator 208 (FIG. 2) generates different models from combinations of the set of features and range of hyperparameters. In some examples, the example query generator 208 generates a plurality of different computer-generated machine learning models based on the plurality of selected combinations of the features and ranges of hyperparameters from the example query selector 206. At block 408, the example query generator 208 triggers parallel training of the models. In some examples, the example query generator 208 initiates the training of the plurality of computer-generated machine learning models based on demographic data of audience measurement panelists that are also subscribers of the database proprietor 102. In the illustrated example, the demographic data is obtained from user accounts of the database proprietor 102 in the privacy-protected cloud environment 106. In some examples, the query generator 208 uses the demographic data of the users from the database proprietor 102, the selected features, and the selected ranges of hyperparameters to train the plurality of computer-generated machine learning models. In some examples, the query generator 208 triggers the training of the computer-generated machine learning models to occur on a computer, server, device, etc. In some examples, the query generator 208 triggers the training of the computer-generated machine learning models by sending a command, instruction, network communication, etc. to the computer, server, device, etc. In some examples, the query generator 208 triggers the training of the computer-generated machine learning models in parallel.

At block 410, the example query generator 208 determines if the models have finished training. If at block 410 the example query generator 208 determines the models have not finished training, the instructions 400 remain at block 410 and wait for the example query generator 208 to determine the models have finished training. If at block 410 the example query generator 208 determines the models have finished training, the instructions 410 continue to block 412 at which the example analytics controller 210 (FIG. 2) generates performance results for all of the models. In some examples, the example analytics controller 210 (FIG. 2) generates performance results for all of the computer-generated machine learning models from the example query generator 208. The example analytics controller 210 compares the results (e.g., demographic information for users, etc.) from the training of the plurality of computer-generated machine learning models to the demographic data from audience measurement panelists from the AME panel data. In some examples, the demographic data from audience measurement panelists from the AME panel data is used to validate (e.g., the AME panel data serves as the truth data) the demographic results of the computer-generated machine learning models. In some examples, the analytics controller 210 obtains demographic data of the audience measurement panelist who accessed media via panelist client devices that correspond to the users of the database proprietor 102 to determine the performance of each of the computer-generated machine learning models. In some examples, the performance results include model accuracy, demographic accuracy, etc.

At block 414, the example analytics controller 210 stores the performance results of the models. In some examples, the example analytics controller 210 stores the computer-generated machine learning models and the corresponding performance results in the example demographic correction models database 142 of FIG. 1. After the example analytics controller 210 stores the performance results of the models, the instructions 400 of FIG. 4 end.

FIG. 5 is a flowchart representative of machine readable instructions 500 which may be executed to implement the example model analyzer 144 of FIGS. 1 and/or 3. In some examples, the instructions 500 are executed by the same computer or machine that executes the instructions 400. In other examples, the instructions 500 are executed by a separate computer or machine than a computer or machine that executes the instructions 400. In this manner, the instructions 400 and the instructions 500 can be flexibly executed by the same computer/machine or by separate computers/machines. The example instructions 500 begin at block 502 at which the example query results interface 302 (FIG. 3) obtains the results of the models. In some examples, the example query results interface 302 obtains the computer-generated machine learning models and corresponding performance results from the example analytics controller 210 of FIG. 2. In some examples, the example query results interface 302 obtains the computer-generated machine learning models and corresponding performance results stored in the example demographic correction models database 142 of FIG. 1.

At block 504, the example data aggregation controller 304 (FIG. 3) runs a query to merge individual results to generate aggregate results. In some examples, the example data aggregation controller 304 generates aggregate results of the performance results from the plurality of computer-generated machine learning models. In some examples, the data aggregation controller 304 aggregates the performance results into a tabular format that is easier to read and interpret when analyzing the performance results of the computer-generated machine learning models.

At block 506, the example model selector 306 (FIG. 3) compares the model output performance results. In some examples, the example model selector 306 compares the performance results of the computer-generated machine learning models based on the aggregated performance results from the example data aggregation controller 304. At block 508, the example model selector 306 selects a model based on the performance results. In some examples, the example model selector 306 compares the performance results across the computer-generated machine learning models to select a computer-generated machine learning model with the best performance relative to other ones of the computer-generated machine learning models. In some examples, the example model selector 306 determines a computer-generated machine learning model with the relatively best performance based on a combination of the performance metrics determined by the example analytics controller 210 of FIG. 2 (e.g., model accuracy, demographic accuracy, etc.). The example model selector 306 determines which combination of features and hyperparameters yielded the best performance results among the computer-generated machine learning models.

At block 510, the example model selector 306 applies the selected model to correct error(s). In some examples, the example model selector 306 applies the selected computer-generated machine learning model to the impressions from the database proprietor 102 to correct for any human errors (e.g., errors from users self-declaring inaccurate demographic information such as, age, gender, etc.) and/or any computer-generated errors (e.g., misattribution errors, non-coverage errors, etc.) in the user demographic information. In some examples, the example model selector 306 uses the selected computer-generated machine learning model to analyze the impressions in the example enriched impressions database 120 of FIG. 1 that were matched to a particular user of the database proprietor 102. The demographic information (e.g., age) from applying the selected computer-generated machine learning for each user may be stored in the example model inferences database 146 of FIG. 1 for subsequent use, retrieval, and/or analysis. In some examples, the example model selector 306 uses the selected computer-generated machine learning model to analyze the entire user base of the database proprietor 102 regardless of whether the users are matched to any particular media impressions. After inferring the correct demographic (e.g., age) for each user using the selected computer-generated machine learning model, the example model selector 306 stores the inferences in the example model inferences database 146 of FIG. 1. In some such examples, when the users matched to particular impressions are to be analyzed (e.g., the users matched to impressions in the enriched impressions database 120), the example model selector 306 extracts the inferred demographic assignment to each relevant user in the enriched impressions database 120 that matches with one or more media impressions. After the example model selector 306 applies the selected model to correct error(s), the instructions 500 of FIG. 5 end.

In some examples, block 510 (e.g., applying the selected model to correct error(s)) can be performed by the same computer that selects the computer-generated machine learning model (as illustrated in FIG. 5). However, in other examples, a separate computer may apply the selected computer-generated machine learning model to correct for the computer-generated errors. For example, one computer may generate the machine learning models and analyze the performance for selected one of the machine learning models, and a separate computer may apply the selected machine learning model to correct the demographic information that may have computer-generated errors.

FIG. 6 is a block diagram of an example processor platform 600 structured to execute the instructions of FIGS. 4 and/or 5 to implement the model generator 140 and the example model analyzer 144 of FIGS. 1-3. The processor platform 600 can be, for example, a server, a personal computer, a workstation, a self-learning machine (e.g., a neural network), a mobile device (e.g., a cell phone, a smart phone, a tablet such as an iPad), a personal digital assistant (PDA), an Internet appliance, a set top box, a headset or other wearable device, or any other type of computing device.

The processor platform 600 of the illustrated example includes a processor 612. The processor 612 of the illustrated example is hardware. For example, the processor 612 can be implemented by one or more integrated circuits, logic circuits, microprocessors, GPUs, DSPs, or controllers from any desired family or manufacturer. The hardware processor may be a semiconductor based (e.g., silicon based) device. In this example, the processor 612 implements the example feature interface 202, the example hyperparameter interface 204, the example query selector 206, the example query generator 208, the example analytics controller 210, the example query results interface 302, the example data aggregation controller 304, and the example model selector 306.

The processor 612 of the illustrated example includes a local memory 613 (e.g., a cache). The processor 612 of the illustrated example is in communication with a main memory including a volatile memory 614 and a non-volatile memory 616 via a bus 618. The volatile memory 614 may be implemented by Synchronous Dynamic Random Access Memory (SDRAM), Dynamic Random Access Memory (DRAM), RAMBUS® Dynamic Random Access Memory (RDRAM®) and/or any other type of random access memory device. The non-volatile memory 616 may be implemented by flash memory and/or any other desired type of memory device. Access to the main memory 614, 616 is controlled by a memory controller.

The processor platform 600 of the illustrated example also includes an interface circuit 620. The interface circuit 620 may be implemented by any type of interface standard, such as an Ethernet interface, a universal serial bus (USB), a Bluetooth® interface, a near field communication (NFC) interface, and/or a PCI express interface.

In the illustrated example, one or more input devices 622 are connected to the interface circuit 620. The input device(s) 622 permit(s) a user to enter data and/or commands into the processor 612. The input device(s) can be implemented by, for example, an audio sensor, a microphone, a camera (still or video), a keyboard, a button, a mouse, a touchscreen, a track-pad, a trackball, isopoint and/or a voice recognition system.

One or more output devices 624 are also connected to the interface circuit 620 of the illustrated example. The output devices 624 can be implemented, for example, by display devices (e.g., a light emitting diode (LED), an organic light emitting diode (OLED), a liquid crystal display (LCD), a cathode ray tube display (CRT), an in-place switching (IPS) display, a touchscreen, etc.), a tactile output device, a printer and/or speaker. The interface circuit 620 of the illustrated example, thus, typically includes a graphics driver card, a graphics driver chip and/or a graphics driver processor.

The interface circuit 620 of the illustrated example also includes a communication device such as a transmitter, a receiver, a transceiver, a modem, a residential gateway, a wireless access point, and/or a network interface to facilitate exchange of data with external machines (e.g., computing devices of any kind) via a network 626. The communication can be via, for example, an Ethernet connection, a digital subscriber line (DSL) connection, a telephone line connection, a coaxial cable system, a satellite system, a line-of-site wireless system, a cellular telephone system, etc.

The processor platform 600 of the illustrated example also includes one or more mass storage devices 628 for storing software and/or data. Examples of such mass storage devices 628 include floppy disk drives, hard drive disks, compact disk drives, Blu-ray disk drives, redundant array of independent disks (RAID) systems, and digital versatile disk (DVD) drives.

Machine executable instructions 632 represented in FIGS. 4 and 5 may be stored in the mass storage device 628, in the volatile memory 614, in the non-volatile memory 616, and/or on a removable non-transitory computer readable storage medium such as a CD or DVD.

From the foregoing, it will be appreciated that example methods, apparatus and articles of manufacture have been disclosed that generate computer-trained machine learning models to correct computer-generated errors in audience data. The disclosed methods, apparatus and articles of manufacture generate a plurality of computer-trained machine learning models with different combination of features and hyperparameters to determine the best performing machine learning model relative to other ones of the machine learning models to correct computer-generated errors in audience data. The disclosed methods, apparatus and articles of manufacture improve the efficiency of using a computing device by running the plurality of machine-trained machine learning models in parallel. In some examples, running the plurality of machine-trained machine learning models manually would take approximately ten minutes for a single machine learning model (e.g., one feature and hyperparameter combination). The disclosed methods, apparatus and articles of manufacture are able to generate and run upward of 100 different machine learning models in ten minutes. The disclosed methods, apparatus and articles of manufacture are accordingly directed to one or more improvement(s) in the functioning of a computer.

Example methods, apparatus, systems, and articles of manufacture to generate computer-trained machine learning models to correct computer-generated errors in audience data are disclosed herein. Further examples and combinations thereof include the following:

Example 1 includes an apparatus comprising a query selector to select a plurality of features and a range of hyperparameters, a query generator to generate a plurality of machine learning models based on the plurality of features and the range of hyperparameters, and initiate training of the plurality of machine learning models based on demographic data in a privacy-protected cloud environment, the demographic data obtained from database proprietor user accounts corresponding to audience measurement panelists, and a model selector to select a first machine learning model from the plurality of machine learning models.

Example 2 includes the apparatus of example 1, wherein the query generator is to initiate the training of the plurality of machine learning models in parallel.

Example 3 includes the apparatus of example 1, further including an analytics controller to generate performance results for the plurality of machine learning models.

Example 4 includes the apparatus of example 3, wherein the analytics controller is to compare results from training the plurality of machine learning models to at least some of the demographic data of ones of the audience measurement panelists who access media via panelist client devices, and generate the performance results based on the comparison.

Example 5 includes the apparatus of example 4, wherein the performance results include at least one of model accuracy or demographic accuracy.

Example 6 includes the apparatus of example 4, further including a data aggregation controller to aggregate the performance results of the plurality of machine learning models.

Example 7 includes the apparatus of example 6, wherein the model selector is to select the first machine learning model from the plurality of machine learning models based on the aggregated performance results.

Example 8 includes the apparatus of example 1, wherein the privacy-protected cloud environment includes first data from at least one of media providers or third parties combined with second data from a database proprietor in a data store, the second data including the demographic data.

Example 9 includes a non-transitory computer readable storage medium comprising instructions that, when executed, cause at least one processor to select a plurality of features and a range of hyperparameters, generate a plurality of machine learning models based on the plurality of features and the range of hyperparameters, initiate training of the plurality of machine learning models based on demographic data in a privacy-protected cloud environment, the demographic data obtained from database proprietor user accounts corresponding to audience measurement panelists, and select a first machine learning model from the plurality of machine learning models.

Example 10 includes the non-transitory computer readable storage medium of example 9, wherein the instructions, when executed, cause the at least one processor to initiate the training of the plurality of machine learning models in parallel.

Example 11 includes the non-transitory computer readable storage medium of example 9, wherein the instructions, when executed, cause the at least one processor to generate performance results for the plurality of machine learning models.

Example 12 includes the non-transitory computer readable storage medium of example 11, wherein the instructions, when executed, cause the at least one processor to compare results from training the plurality of machine learning models to at least some of the demographic data of ones of the audience measurement panelists who access media via panelist client devices, and generate the performance results based on the comparison.

Example 13 includes the non-transitory computer readable storage medium of example 12, wherein the performance results include at least one of model accuracy or demographic accuracy.

Example 14 includes the non-transitory computer readable storage medium of example 12, wherein the instructions, when executed, cause the at least one processor to aggregate the performance results of the plurality of machine learning models.

Example 15 includes the non-transitory computer readable storage medium of example 14, wherein the instructions, when executed, cause the at least one processor to select the first machine learning model from the plurality of machine learning models based on the aggregated performance results.

Example 16 includes the non-transitory computer readable storage medium of example 9, wherein the privacy-protected cloud environment includes first data from at least one of media providers or third parties combined with second data from a database proprietor in a data store, the second data including the demographic data.

Example 17 includes a method comprising selecting a plurality of features and a range of hyperparameters, generating a plurality of machine learning models based on the plurality of features and the range of hyperparameters, initiating training of the plurality of machine learning models based on demographic data in a privacy-protected cloud environment, the demographic data obtained from database proprietor user accounts corresponding to audience measurement panelists, and selecting a first machine learning model from the plurality of machine learning models.

Example 18 includes the method of example 17, further including initiating the training of the plurality of machine learning models in parallel.

Example 19 includes the method of example 17, further including generating performance results for the plurality of machine learning models.

Example 20 includes the method of example 19, further including generating the performance results by comparing results from training the plurality of machine learning models to at least some of the demographic data of ones of the audience measurement panelists who access media via panelist client devices, and generating the performance results based on the comparison.

Example 21 includes the method of example 20, wherein the performance results include at least one of model accuracy or demographic accuracy.

Example 22 includes the method of example 20, further including aggregating the performance results of the plurality of machine learning models.

Example 23 includes the method of example 22, further including selecting the first machine learning model from the plurality of machine learning models based on the aggregated performance results.

Example 24 includes the method of example 17, wherein the privacy-protected cloud environment includes first data from at least one of media providers or third parties combined with second data from a database proprietor in a data store, the second data including the demographic data.

Example 25 includes an apparatus comprising memory, and at least one processor to execute computer readable instructions to at least select a plurality of features and a range of hyperparameters, generate a plurality of machine learning models based on the plurality of features and the range of hyperparameters, initiate training of the plurality of machine learning models based on demographic data in a privacy-protected cloud environment, the demographic data obtained from database proprietor user accounts corresponding to audience measurement panelists, and select a first machine learning model from the plurality of machine learning models.

Example 26 includes the apparatus of example 25, wherein the at least one processor is to execute the computer readable instructions to initiate the training of the plurality of machine learning models in parallel.

Example 27 includes the apparatus of example 25, wherein the at least one processor is to execute the computer readable instructions to generate performance results for the plurality of machine learning models.

Example 28 includes the apparatus of example 27, wherein the at least one processor is to execute the computer readable instructions to compare results from training the plurality of machine learning models to at least some of the demographic data of ones of the audience measurement panelists who access media via panelist client devices, and generate the performance results based on the comparison.

Example 29 includes the apparatus of example 28, wherein the performance results include at least one of model accuracy or demographic accuracy.

Example 30 includes the apparatus of example 28, wherein the at least one processor is to execute the computer readable instructions to aggregate the performance results of the plurality of machine learning models.

Example 31 includes the apparatus of example 30, wherein the at least one processor is to execute the computer readable instructions to select the first machine learning model from the plurality of machine learning models based on the aggregated performance results.

Example 32 includes the apparatus of example 25, wherein the privacy-protected cloud environment includes first data from at least one of media providers or third parties combined with second data from a database proprietor in a data store, the second data including the demographic data.

Although certain example methods, apparatus and articles of manufacture have been disclosed herein, the scope of coverage of this patent is not limited thereto. On the contrary, this patent covers all methods, apparatus and articles of manufacture fairly falling within the scope of the claims of this patent.

METHODS AND APPARATUS TO GENERATE COMPUTER-TRAINED MACHINE LEARNING MODELS TO CORRECT COMPUTER-GENERATED ERRORS IN AUDIENCE DATA

Information

Publication Number

Date Filed

Date Published

Inventors

Original Assignees

CPC

International Classifications

Abstract

Description

Claims

RELATED APPLICATION(S)

Provisional Applications (1)