Patent Grant
9918142
The present disclosure relates to systems and methods for evaluating television media instances for advertisement spots based on various factors for reaching television viewers who are desired product buyers.
Television is the largest advertising medium in the United States, with over 65 billion dollars in advertising revenue in 2011. According to Nielsen, approximately 20 times more hours are spent viewing TV as compared to viewings on either the Internet or mobile video. In 2013, there were about twice as many original programs on TV as compared to 2005, and over 60% of viewers were using High Definition (“HD”) TVs.
If there is an area for improvement in TV, it is around how advertising can be effective and targeted to viewers. TV advertising is unlike online advertising because it has traditionally been a broadcast medium, i.e., a one way transmission of TV programs to the viewer with no direct feedback. In online advertising, it is possible to deliver ads to individual persons, via cookies and IP addresses, and to then track the behavior of those persons, including whether they convert after seeing the advertisement by observing their clicks on advertisements and conversions on web sites.
In TV, advertisements may be embedded in a single high definition video stream, and broadcast using over-the-air terrestrial transmission towers, satellite, and/or cable. The single signal transmission enables high bandwidth and very high quality TV signal. However, this introduces significant limitations. Apart from small experimental TV systems, there are currently no available technologies for delivering advertisements one-to-one to households at a scale equivalent to TV broadcasting.
A second major limitation is determining whether a purchase was influenced by the TV advertisement. Standard TV systems do not allow advertisers to know if individuals saw the advertisements. Further, standard TV systems cannot determine if an individual who is purchasing a product or service, saw the advertisement.
Because of these and other limitations, since the 1950s, this medium has been tracked using a 25,000 person, Nielsen “panel” with “diaries.” The individuals on Nielsen's panel could report on what they saw on TV, and then this data could be extrapolated across the United States (115,000,000 households). This panel is both small and yet expensive to maintain. However, in the United States, set top boxes (“STBs”) are now present in over 91.5% of US homes. Further, since 2009, STBs with return path capabilities have proliferated in the United States, comprising over 30% of STBs in households. The number of households with STBs is greater in size than the Nielsen panel, and the scale and richness of detail of STB data allows for new capabilities in TV advertisement targeting.
In order to utilize new capabilities, the present disclosure relates to systems and methods that use current U.S. data collection and U.S. TV broadcasting capabilities. As will be discussed in further detail below, the systems and methods discussed herein provide a framework for understanding certain TV targeting problems and approaches for solving them. Benefits of the present disclosure may include providing detailed descriptions of data formats available for television targeting; formalizing TV advertisement targeting problems into one or more objective functions; identifying variables available for advertisement targeting that can be used for targeting practical TV advertisement campaigns; providing a plurality of algorithms for TV data; and combining the plurality of algorithms to provide desired results.
According to certain embodiments, methods are disclosed for teaching a television targeting system to reach product buyers. One method includes receiving, at a server, one or more heterogeneous sources of media data, the media data including television viewing events; generating, by the server, a plurality of media asset patterns from the one or more heterogeneous sources of media data, the plurality of media asset patterns being possible media placements which are represented as conjunctive expressions; calculating, by the server, one or more heterogeneous advertisement effectiveness measures for each media asset pattern; calculating, by the server for a plurality of pairs of an advertisement and a media instance, a number of previously placed airings of the advertisement in the media instance; and generating, by the server, a model to predict advertisement effectiveness for each pair of an advertisement and a media instance based on a combination of the ad effectiveness measures and the number of previously placed airings of the advertisement in the media instance.
According to certain embodiments, systems are disclosed for teaching a television targeting system to reach product buyers. One system includes a data storage device storing instructions; and a processor configured to execute the instructions to perform a method including: receiving, at a server, one or more heterogeneous sources of media data, the media data including television viewing events; generating, by the server, a plurality of media asset patterns from the one or more heterogeneous sources of media data, the plurality of media asset patterns being possible media placements which are represented as conjunctive expressions; calculating, by the server, one or more heterogeneous advertisement effectiveness measures for each media asset pattern; calculating, by the server for a plurality of pairs of an advertisement and a media instance, a number of previously placed airings of the advertisement in the media instance; and generating, by the server, a model to predict advertisement effectiveness for each pairing of an advertisement and a media instance based on a combination of the ad effectiveness measures and the number of previously placed airings of the advertisement in the media instance.
Additional objects and advantages of the disclosed embodiments will be set forth in part in the description that follows, and in part will be apparent from the description, or may be learned by practice of the disclosed embodiments. The objects and advantages of the disclosed embodiments will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims. As will be apparent from the embodiments below, an advantage to the disclosed systems and methods is that multiple parties may fully utilize their data without allowing others to have direct access to raw data. The disclosed systems and methods discussed below may allow advertisers to understand users' online behaviors through the indirect use of raw data and may maintain privacy of the users and the data.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosed embodiments, as claimed.
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate various exemplary embodiments and together with the description, serve to explain the principles of the disclosed embodiments.
Aspects of the present disclosure, as described herein, relate to determining what television programs to place advertisements on for certain products, by evaluating aspects of the viewers of those television programs. Aspects of the present disclosure involve recognizing that media may be represented and evaluated by the demographics of the people who watch that media. The system may perform a match against media by looking for the television program whose viewers are the closest match to the customers that buy the product to be advertised. After the system finds a close match, it may recommend buying that media (i.e., placing the product ad within that television program). Aspects of the present disclosure may use targeting capabilities, tracking, and delivery, and may add in individualized information to its demographic segment information in order to improve the matching quality.
In one embodiment, the method used by a media buyer may include using Nielsen aggregated data to determine which program to purchase. Furthermore, while a Nielsen panel may be a useful data source and use of this data is described in this disclosure, the Nielsen viewer panel may be somewhat limited by its relatively small size, and limitations in covering certain geographic areas. Accordingly, a variety of enhancements are discussed for making the techniques described below compatible with multiple other data sources (including census data, set top box data, and linked buyer data) so as to create a highly complete and rich profile based on millions of viewers, over 400 variables, and buyers rather than viewers.
Various examples of the present disclosure will now be described. The following description provides specific details for a thorough understanding and enabling description of these examples. One skilled in the relevant art will understand, however, that the present disclosure may be practiced without many of these details. Likewise, one skilled in the relevant art will also understand that the present disclosure may include many other related features not described in detail herein. Additionally, some understood structures or functions may not be shown or described in detail below, so as to avoid unnecessarily obscuring the relevant description.
The terminology used below may be interpreted in its broadest reasonable manner, even though it is being used in conjunction with a detailed description of certain specific examples of the present disclosure. Indeed, certain terms may even be emphasized below; however, any terminology intended to be interpreted in any restricted manner will be overtly and specifically defined as such in this Detailed Description section.
The systems and method of the present disclosure allow for the receiving and processing of TV (media) related data and consumer related data from a plurality of different data sources and of a variety of different data types and formats. Based on the received data, the systems and methods may build a model that may be used to estimate a probability of reaching a particular set of persons. The estimated probability may then be used to determine a value associated with buying an advertisement spot within a television program for the advertisement.
System Architecture
Any suitable system infrastructure may be put into place to receive media related data to develop a model for targeted advertising for television media.
Aspects of the present disclosure may be embodied in a special purpose computer and/or data processor that is specifically programmed, configured, and/or constructed to perform one or more of the computer-executable instructions explained in detail herein. While aspects of the present disclosure, such as certain functions, are described as being performed exclusively on a single device, the present disclosure may also be practiced in distributed environments where functions or modules are shared among disparate processing devices, which are linked through a communications network, such as a Local Area Network (“LAN”), Wide Area Network (“WAN”), and/or the Internet. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
Aspects of the present disclosure may be stored and/or distributed on non-transitory computer-readable media, including magnetically or optically readable computer discs, hard-wired or preprogrammed chips (e.g., EEPROM semiconductor chips), nanotechnology memory, biological memory, or other data storage media. Alternatively, computer implemented instructions, data structures, screen displays, and other data under aspects of the present disclosure may be distributed over the Internet and/or over other networks (including wireless networks), on a propagated signal on a propagation medium (e.g., an electromagnetic wave(s), a sound wave, etc.) over a period of time, and/or they may be provided on any analog or digital network (packet switched, circuit switched, or other scheme).
Use of the system of
One step may be to setup data feeds with one or more media agencies, which may ensure the collection of all the data about what media is being purchased, running, and trafficked to stations. This may also ensure that there is an accurate representation of the available television media. This step may include setting up data feeds for one or more of: media plan data (e.g., as shown in
Media plan data may include a station a commercial will run on, an advertiser, topic information, a media cost associated with the purchase, a phone number, and/or a web address that is associated with the commercial for tracking purposes.
Third-party verification services may watermark commercials and monitor when the media was run across all TV stations. The data generated by third-party verification services may be used to verify that a media instance that was purchased for an advertisement spot was actually displayed on TV.
The sample trafficking instructions and/or order confirmation may include a product that was purchased, and instructions that a station is to use when displaying a commercial.
Another step may be to setup data feeds with one or more call centers, which may ensure there is accurate data about callers that called into specific phone numbers. This step may include receiving a call center data feed (e.g., as shown in
Yet another step may be to setup one or more data e-commerce vendor data feeds. E-commerce data feeds may be setup to receive recurring data feeds with a vendor and/or internal system of an advertiser that records orders that come in from an advertiser's website (e.g., as shown in
Another step may be to setup one or more data order processing/fulfillment data feeds. Data order processing/fulfillment data feeds may be setup to receive recurring data feeds with order vendor and/or internal system that physically handles the logistics of billing and/or fulfillment. This step may ensure an accounting of subsequent purchases, such as subscriptions and for returns/bad debt, etc., and may ensure accurate accounting for revenue. This step may also include receiving data from a series of retail Point of Sale (“PoS”) systems (e.g., as shown in
Another step may be to setup one or more audience data enrichment data feeds with one or more data bureaus. This step may ensure that callers, web-converters, and/or ultimate purchasers have their data attributes appended to their record in terms of demographics, psychographics, behavior, etc. (e.g., as shown in
Yet another step may be to setup one or more data feeds with one or more guide services. This step may ensure that forward looking guide service data is ingested into the system. This data may be programming based on what is going to run on television for the weeks ahead (e.g., as shown in
Another step may be to setup one or more data feeds for panel data enrichment. Data related to purchasers of products on television, set top box viewer records, and/or existing panels may be received as a data feed and appended to an advertiser's purchaser data mentioned above (
In another step, all of the underlying data may be put into production. For example, all of the data feeds setup from steps one through seven may be loaded into an intermediate format for cleansing, adding identifiers, etc. Personally Identifiable Information (“PII”) may also be split and routed to a separate pipeline for secure storage. As shown in
At the next step, media plan data 104a, verification data 104b, and/or trafficking data 104c of the agency data system 104 may be received at a data feed repository 112 of the media processing system 102. Further, call center data 106a, e-commerce data 106b, and/or order management data 106c of advertiser data system 106 may be received at the data feed repository 112. Additionally, viewer panel data 108a, guide data 108b, and/or consumer enrichment data 108c of the audience data system 108 may be received at the data feed repository 112. After one or more of data feeds are received by the feed repository 112, data may be extracted from the data feeds by extractor 114 of media processing system 102.
At another step, business logic/models may be run for matching responses and orders to media (“attribution”). In this step, the data extracted from the data feeds has been ingested into the system at the most granular form. Here, the phone responses may be matched up to media that generated it. The e-commerce orders may be matched using statistical models to the media that likely generated them. As shown in
At yet another step, the analyzed data may be loaded into databases. For example, the data may have already been aggregated and/or final validation of the results may have been completed. After this, the data may be loaded by loader 122 into one or more databases 124 for use with any of the upstream media systems, such as data consumers system 110. These include the ability to support media planning through purchase suggestions, revenue predictions, pricing suggestions, performance results, etc. One or more databases 124 may include customers database 124, campaign database 124, station inventory database 124, performance database 124, models database 124, and/or PII database 124.
At another step, the analyzed data may be used by presentation module 126. In this step, all of the data may be accessible to the operators of various roles in the media lifecycle. This may include graphical tools for media planning (where the targeting in this application primarily fits), optimization, billing, trafficking, reporting, etc.
The above-described system may be used to gather, process, and analyze TV related data. This data may then be used to identify certain available media instances, or advertisement spots, that an advertiser may purchase to display an advertisement. As will be described in further detail below, advertisement spots, also referred to as media instances, may be evaluated and scored to assist an advertiser in choosing which media instance to purchase.
Media Instances
As described above, a TV media instance, Mi, may be any segment of time on TV that may be purchased for advertising. The media instance, Mi, as an element of the Cartesian product, may be defined as follows:
MiεS×P×D×H×T×G×POD×POS×L
where S is station, P is program, D is day-of-week, H is hour-of-day, T is calendar-time, G is geography, POD is the ad-pod, POS is the pod-position, and L is media-length.
Stations may include broadcast and/or cable stations, and may be identified by their respective call letters, such as KIRO and CNN. Geography may include national (nationwide), one or more direct market association areas, such as Miami, Fla., and/or cable zones, such as Comcast Miami Beach.
An ad-pod may be a set of advertisements that run contiguously in time during a commercial break for a TV program. Pod-position may be the sequential order of the advertisement within its pod. Media length may be the duration of the time segment in seconds. Media length, for example, may include 15, 30, 45, and/or 60 second spots.
The present disclosure allows the advertiser to select a set of media instances, Mi, to purchase for advertisement targeting for an ideal audience. The present disclosure also allows the advertiser to provide a bid, CPI(Mi) cost per impression, such that the expected advertisement response per dollar is maximized, as follows:
Mi: max ΣirpiΩ(Mi)·I(Mi)
subject to ΣiCPI(Mi)·l(Mi)≦B and V({Mi})=true
where rpiΩ(Mi) is the response (also referred to as a conversion, a sale, and/or revenue) per impression or target-audience-concentration per impression or probability-of-target-audience per impression for the given media instance, Mi; I(Mi) are the impressions for media instance, Mi; B is the TV campaign budget; and V determines if the set of media instances, Mi, violates advertiser-defined rotation rules. Rotation rules may be, for example, running an advertisement no more than once per 60 minutes, having no greater than 5% of budget on any one network or day-part, etc. Rotation rules may be defined by TV advertisement buyers and/or broadcast networks.
One embodiment of the present disclosure is to iteratively select media instances in order of value per dollar, as follows:
subject to rotation rule constraints V until the budget is filled. CPI(Mi) and rpiΩ(Mi) are both estimates using historical clearing prices and media observations.
Methods will next be described for estimating the response per impression or target-audience-concentration per impression “rpiΩ(Mi)” part of the formula above.
Media Asset Patterns
A media asset pattern may be any set of variable value instantiations of a media instance. Formally, media asset pattern, may be a subset of instantiated features from the media instance Mi mi,t⊂Mi, for example, a future media instance that is under consideration to buy may be Mi=(CNN, 8 pm, “Piers Morgan”, Tuesday, 12/12/2012, Pod1, Pos2, 60 s). The following media asset patterns may be used to predict its performance: Station mi1=(CNN); Station-Hour-Pod mi2=(CNN, 8 pm, Pod1); Geography-Station mi3=(National-CNN); and others.
Table 1, below, shows a list of Media Asset Patterns used in one embodiment of the present disclosure.
Examples of various media asset pattern types will now be described in more detail.
TV programs are intuitively what people tune into when watching television. Different programs appeal to different people. For example, viewers of TLC's “I Didn't Know I Was Pregnant” may be different from viewers of SYFY's “Continuum.”
There are over 450,000 weekpart-daypart-programs available to be purchased on TV. The programs may be good predictors of advertisement performance. An example of media asset patterns and their calculated ad effectiveness scores is shown in table 2, below.
In addition to using programs in general, it is also possible to demarcate a special class of programs which may be referred to as “high impact programs.” These programs have high observed impressions per expected impressions for their station-timeslot,
Impactful programs may include event programs, such as “The Academy Awards,” football games, and very popular reality programs, e.g., “Dancing with the Stars.” Impactful programs may also include “cultural phenomena,” such as “Honey Badgers!” Table 3 below depicts programs and their respective impressions performance relative to their expected performance in their timeslot. A media asset pattern of High-Impact-Program can then be established and used by the system.
Stations often run similar programming in the same station-day-hour timeslots. This information may add value as a predictor, as some demographics may have a propensity to watch TV on certain times of day. For example, high income people tend to watch in prime-time, but not daytime. Weekday, daytime programming may be highly skewed toward older and/or lower income households.
Program names are often recorded in television panel data and schedules in a variety of inconsistent ways, often because television program data is hand-entered. Thus when a buyer is attempting to buy “Cold Case,” the present disclosure may fail to find a match for “Cold Case” because the panel data might have recorded this as “Cold Case Sat.” In order to address this, the presently disclosed methods may use a series of mapping tables to map native panel strings to “mastered” versions of those strings, which may facilitate matching. The present disclosure also allows editors to inspect the native strings, and uses edit distance to identify similar mastered strings that each native string may be mapped to. These “mastered” program names are then used in media asset patterns. Examples of program master mappings are shown in Tables 5A-5C, below.
Table 6, below, depicts an exemplary MediaAssetPatternType 53-AgeGender-Station-Program showing entries like “Academy Award”. Note that these programs are all actually the same program. “LRC” stands for “Live from the Red Carpet.”
Table 7, below, depicts an exemplary MAPType 59-AgeGender-Station-ProgramMaster showing programs like “Academy Awards”. The various program strings have been remapped to a single canonical program called “Live from the Red Carpet: The Academy Award.”
Human viewing behavior is periodic and so viewers of a program this week are likely to have also viewed the same program in the previous week. TV Program episodes are often sequential in that the story builds from one week to the next, or sports games follow events from the previous week, and in the same way, human viewing tends to track the episodes from week to week. During some seasons, viewership increases from episode to episode (e.g., see
Table 8, below, depicts an exemplary Media Asset Pattern Type 98-Station-Program-Hour-Prior 1 week: This shows the impressions generated by the same program, at the same hour 1 week prior to the airing.
Premiere or First-time-Airings of Episodes for programs such as The Walking Dead tend to attract large viewing audiences. These premieres are often followed by a “same-day encore,” and then some repeats during the week. The audiences are much smaller for repeats that were first shown 40 or 100 days ago.
This phenomenon may be used to create a time-since-first-airing media asset pattern. This is a number 0 or higher (or coded as Station-Program-first-day, Station-Program-first-day encore, Station-Program-first-week, Station-Program-more than 1 week) which can be used to predict the audience and viewing audience impressions given a certain time since the first airing. In order to calculate this, the first detected episode number may be used to take the date of the first airing, and then take the fractional number of days since the first detection.
Table 9, below, depicts an exemplary Media Asset Pattern Type for Time-Since-First-Airing: Times are discretized into 0 (premiere), 0.5 (same-day encore), 7 (same week) and 8 (greater than 1 week since the first detected airing).
The above shows how viewership changes fairly dramatically with the premiere, encore, or repeat. A feature called time-since-last-airing may be used to help to predict the viewership of each program. The time-since-first-airing starts at the far left with a high value indicating that these are re-runs from last year. Then when the premiere is shown, the time-since-last-first-airing drops to 0 and there is a spike on viewing. After that it may be possible to see that time-since-first-airing changes between 0 and 7, and that the associated changes in viewing may be seen.
Pod position and commercial break are also important features of the ad insertion, and can dramatically affect the viewership and response per impression from the ad. In general the first pod has the highest viewership, and viewership then decreases throughout the commercial break.
Using the above pod information it is possible to create media asset patterns of the form: Station-Program-PodSequence and to estimate performance of these differentially.
TV broadcasts can be performed nationally and locally. Advertisers often execute local TV campaigns when they are trying to get very precise levels of targeting, for example during elections. Often particular geographic markets such as Birmingham Ala. behave differently to overall national population. For example, Montel may over-perform—have more engaged viewers—in the South and under-perform in the North. It may be possible to represent media as Market-Station-Day-Hour or Market-Station-Program and then measure the ad effectiveness or response per impression from these different markets, and use these in an ad targeting system.
Because there are a large number of local markets (210 DMAs), it is desirable to control the amount of data being retained. One embodiment utilizes a feature whereby it calculates the RPI or ad effectiveness metrics for each market, and then if the RPI metric is not significantly different (as measured in absolute difference) from the national RPI metric or ad effectiveness metric, then the local ad effectiveness metric can be deleted (converted to missing), which as described below, may result in the national RPI or ad effectiveness metric being used. The degree of absolute difference is a parameter that can be used to control how much local data is retained.
Table 10, below, depicts an example Market-Station-Program media asset patterns for a range of geographies and the same program. This shows that the estimated ad effectiveness varies by geography.
Table 11, below, depicts an example Market-Station-Program media asset patterns and their ad effectiveness scores. The market shown is Birmingham, Ala.
Viewership changes throughout the year, in some part in response to programming changes, but in other parts due to different events that occur each year. for example, each December, Hallmark's viewership increases dramatically as they air a variety of family favorite Christmas movies.
As shown in
Media Assets can also be represented by their Genre. Table 13, below, shows genres as classified by Nielsen corporation using their taxonomy, and how programs in those genres were scored for a demographic match to buyers. For example, Devotional is the genre that has the highest correlation with buyers—a result which makes sense as these customers tend to be religious and view a lot of religious programming.
TV broadcasts occur locally and nationally. It may even be possible to use data about the sales per capita in particular geographic areas to inform the presently disclosed system as to the expected response from these areas when an ad is broadcast in these areas. The media asset pattern type in this case is simply a local market which may or may not include the program information.
Media Asset Patterns can also be represented by the keywords of program names. An example is shown in table 14, below. When the keywords below are in the program title, impressions are on average higher than expected. It is possible to create Media Asset Patterns for Genre-keyword.
Advertisement Response:
Advertisement response is a generalized measure of the concentration of a desired audience within a particular media asset pattern Mi. This may be calculated using several measures including the number of buyers reached by targeting each media asset, phone response per impression, the concentration of targeted audience, and others. In one embodiment, information about response may come from any subsystems of data feeds of advertiser data system 106.
Advertisement response may be represented as RΩ(
RΩ(
TV Advertisement Response/Ad Effectiveness Measures:
Multiple ad effectiveness measures may be used for helping to estimate response per impression or concentration of target audience per impression. One method may be Target Rating Points (“TRPs”) on Age-Gender.
Target Rating Points (TRPs) on Age-Gender:
Age-gender Target Rating Points may be used as a form of targeting. This form of targeting may be based on the number of persons who match the advertiser's target demographics divided by total viewing persons. A formula representing age-gender TRPs may be represented as:
where Q(mi) is a set of viewers who are watching TV media instance mi; where this viewing activity was recorded by Nielsen panel; where qkεQ(mi); where # is the cardinality of a set; and where #rT includes persons that match on all demographics.
For example, a calculation of rA(P, mi) as 50% may mean that 50% of the people are a match to the desired demographics. Age-gender TRPs may also be calculated using Nielsen “Market Breaks,” such as gender=male|female and/or age=18-24, 25-34, 35-44, 45-54, 55-64, 65+.
Table 15, below, depicts an example of MAPType 59 with Ad Effectiveness of Target Rating Points (TRPs).
Phone Response Per Impression:
When a TV advertisement is run with a 1800 number, it may be possible to match the phone responses on specific 1800 numbers back to the advertisement that was placed. This data may be used to track sales due to the TV advertisement. A specific method may use a series of hour lag terms to predict the number of phone-calls that would be generated on a given hour.
The method of the present disclosure exposes hour and day-lag terms for historical phone response, and then trains a system to predict a probability of phone response from an upcoming media spot. The method of the present disclosure may be represented by the formulas:
where CALL(mj,T) are the number of calls from airing mj,T.
Table 16, below, depicts an example of: Media Asset Pattern Type 38-Station-Day-Hour with Ad Effectiveness equal to Phone Responses Per Impression for a Life Insurance product, including a selection of scores for CNN.
Table 17, below, depicts an example of a Media Asset Pattern Type 38-Station-Day-Hour with Ad Effectiveness equal to Phone Responses Per Impression for a Life Insurance product. Scores ordered by RPI descending.
Buyer Ratings:
Buyer targeting may look for media that has a high rate of observed buyers per impression, and targets those programs. An algorithm that may not be trained by itself, such as a self-learning algorithm and/or recursive algorithm, may score a percent of buyers observed in each media, which may be referred to as “buyer ratings.” The following expression defines buyer ratings.
Table 18, below, depicts an example of a Media Asset Pattern Type 47-Station-Program Buyers per impression in the audience (SourceViewPct).
Table 19, below, depicts an example of a Media Asset Pattern Type 47-Station-Program Buyers per impression, sorted in order of highest buyers per impression programs to lowest for Life Insurance Product. A variety of religious programs show up as having high buyers per impression.
High Dimensional Demographic Matching:
In one embodiment, demographic match across 3,000 variables between an ad product buyer and each media asset pattern may also be used. Similar to age-gender matching, demographic mapping may use a thousand times more variables and a different match calculation due to the high dimensionality. The demographic match between an ad product and media may be defined as follows:
where P is a vector of demographics representing the average buyer demographic readings, and M is a vector of demographics for the media placement.
Table 20, below, depicts an example of a Media Asset Pattern Type 24-Station-Program with Ad Effectiveness=High Dimensional Demographic Match between Buyers and Set Top Box Viewers of Program. Selection for DIY channel.
Table 21, below, depicts an example of a Media Asset Pattern Type 24-Station-Program with Ad Effectiveness=High Dimensional Demographic Match between Buyers and Set Top Box Viewers of Program. Top several programs by correlation for a Life Insurance product.
Web Spike Per Impression:
If TV broadcasts are aligned in time and geography with web traffic, the difference in web visits due to each broadcast may be calculated by comparing web activity a few minutes before and after the broadcast. These web spike effects may be highest within about 1 minute to about 5 minutes of an airing. Details on calculation of web spike per impression may be as follows:
where ΔW(mj,T)=W(mj,T,t,g)−W(mj,T,t,g) is the difference in web activity at time t2 vs t1, from the same geographic area.
Table 22, below, depicts an example of a Media Asset Pattern 69 Station-Day-Hour with Ad Effectiveness measure equal to Web Spike Response per impression. Table below is sorted in order of highest web spike response per impression to lowest for a different advertising product (identified by sourcekey=110401). This is a product that appeals to women 25-34. The top networks showing up for webspike response are Soap (SOAP), Comedy (Com), Discovery Health and Fitness (DFH).
The ad targeting algorithms, as shown below, may be a combination of one or more of: (i) ad effectiveness metric; and (ii) media asset pattern type. For example, stbheadmatch-station-day-hour may mean high dimensional match with set top box data using statistics on station-day-hours (e.g., CNN-Tues-8 pm's demographic match between target and viewing audience).
Table 23, bellows, shows the correlation between each ad effectiveness measure and a particular response per impression measure. For example, Media Asset Pattern Type 32-STBHead-Station-Day-Hour has a high correlation with buyers per million (0.8471) and is present 93.9% of the time.
Properties of TV Ad Targeting Algorithms:
One element affecting an ad effective metric's ability to be used may be their sparsity. The most sparse data may be STB buyer data, which may be known persons who have bought the advertiser's product, and are also detected watching a particular program. The probability of detection of these customers may be small.
One key reason for sparsity may be because each person must be matched in both STB data and advertiser data.
High dimensional demographic matching may not be as impacted by sparsity because it may aggregate all STB data into a demographic vector, and then may match using this vector. By converting to a demographic vector, it may be possible to eliminate the need for “cross-domain” person-to-person linkage.
As shown in
Demographic matching may have a shallower slope, as shown in
In terms of usable predictions (scoring airings with impressions such that prediction performance is above 0), in one exemplary, non-limiting embodiment, demographic match covered, e.g., 99% of all airings, RPI covered, e.g., 57% of all airings, and BPI covered, e.g., only 0.5% of all airings, as shown in
Exemplary Robust Algorithm:
One benefit of the present disclosure is that the below described targeting algorithm is able to use all of the above-described data and methods which allows for a “hyper-targeted” TV campaign. In order to build a combined algorithm, various problems introduced by the different metrics and range of each algorithm may be overcome. Further, the combined algorithm may be able to select features that are most predictive, and may be trained.
Model:
In one embodiment, a model consistent with the present disclosure may receive all of the available media asset patterns mi,t and ad effectiveness measures ra(mi,t). The model may also use them to predict the ad response per impression RΩ(Mi). This may include a supervised learning problem, as ad effectiveness information may be available for every airing, and thus, the system may be trained to predict the quantity based on historical examples. The model of the present disclosure may include a stacked estimator where each ad effectiveness model rα(mi,t) is an expert, and the assembly is trained to predict ad response RΩ(Mi).
RΩ(Mi)=Z−1(y,μΩ,σΩ)
The predictors xt and ad response target y=Z(RΩ, μ, σ) may be standardized, as discussed below. In order to handle so many different variables, the model may be able to standardize the different variable and may select the variables that are most useful for predicting its target to avoid over-fitting.
Variable Standardization:
In one embodiment, different ad effectiveness variables, such as telephone response per impression (RPI), buyers per impression (BPI), and demographic match, may be used. Each of these variables may have a different set of units. In order to handle these different scales, variables may be transformed, as follows:
xt=Z(rt); y=Z(RΩ); Z(a)=(a−μ)/σ
When training the system to predict standardized target y for each ad effectiveness predictor xt, each predictor may be effectively measuring the relationship between a change of a unit standard deviation in its distribution, to what that translates into in terms of standard deviations of movement in the target variable. This may have several useful properties, such as no constant terms, interpretability, and/or usability.
A constant term may be in effect removed and the co-variance may be measured. The constant term may be “added back” later when the prediction is converted back into target unit. Interpretability may allow standardizing variables on the same scale. When estimating weights, weights in order of magnitude may be read off, and thus, variables that are contributing most to the prediction may be seen. Usability may allow users to enter their own weights if they have some domain knowledge. Because of standardization, w=0.4 intuitively means that 40% of the decision may be based on this variable.
Constraints Due to Ad Theory:
There may be certain constraints that may be imposed on the model due to experimental findings from advertising theory. Ad theory suggests that as the traits of the ad match the product more, response to advertising should increase. Thus, the following propositions for ad effectiveness metrics may be: (1) ad effectiveness ∀i: xiy>0 (since each ad effectiveness metric may be positively correlated with ad response); and (2) given a model predicting ad response y=Σwtxt ∀t: wt≧0, the effect of improved ad effectiveness may be zero or positive on ad response.
Minimum Weight Constraints:
In order to be consistent with the above-mentioned propositions, a positivity constraint in weights may be:
wt≧0
Sum of Weight Constraints:
For reasons of robustness in production, it may be desired to ensure that predictions do not extrapolate higher or lower than a range of values that has been previously observed. For example, a weight of 2 may allow the system to predict outside of the range of the ad response variable. To ensure the sum of weight constraints, all weights may sum to 1. As a result of this additional constraint, a formula may be:
1≧wt≧0Σt=1Twt=1 (2)
Low Data Behavior/Variable Participation Thresholds:
Each media asset pattern may cover a certain number of historical airings. For each media asset pattern, m, the sum the number of impressions observed may be I(m). Accordingly, the ad effectiveness measures may be unreliable on small amounts of data. Bayesian priors may be used to “fill-in” performance when there is less information available, modifying the ad effectiveness score as follows:
r=e−α·I(m)·r+(1−e−α·I(m))·rPRIOR
where α is a parameter that governs how many impressions are collected for the posterior estimate to be favored more heavily than the prior.
However, Bayesian priors may be incorrect and may involve creation themselves. Since there may be hundreds of thousands of variables per product (not to mention hundreds of products), a large number of parameters may be set. Thus, an effect of poorly set priors may be significant as they cause variables that may have been good predictors to be spoiled, and the training process to be unable to weight them properly.
The system of the present disclosure may be able to work reliably with minimal human intervention. Thus, the system may be trained using participation thresholds. IMIN may be defined as the minimum impressions allowed on a particular media asset pattern. If a media asset pattern fails to meet this threshold, it may be converted to a missing value, and thus, does not participate further. The prediction formula for handling missing values may be defined as:
if I(σt=0 then wt=0;xt=MV
Missing Value Handling:
In certain embodiments, a particular media asset pattern may be missing and/or otherwise may be unable to report a value. For example, a system may not have enough data on a program to be able to provide a prediction. When this happens, the system may use a more general media asset pattern type, such as the station, to provide a prediction. Missing value handling may allow the system to operate in cases where a variable is not available and/or a variable is zeroed out, and missing value handling may allow other variables that are present to be used to create a prediction.
For production robustness, media asset pattern types may be defined with small weights, so that if there is a failure then the system may default to one of these more general media asset pattern types. For example, if station-day-hour is undefined, then station may be defined but at a very low weight. Thus, a significant weight may not be given to missing values.
Transforming into Target Units:
The standardized predictions may be converted into the original units. This may be performed by inverting the z-score transform
Z−1 [y]=yσj+μj
where j is the ad effectiveness measure that is being reported. The Z−1 transform may be similar to performing a programming language cast operation into the appropriate units.
Training Algorithm:
Weight training may use a subspace trust-region method that is designed to operation for values 0 to 1 and sum of weights=1 constraints, as shown below:
A forward-backward selection algorithm may be used to select new features to include in the model.
Different Target Value Types:
The Scoring Service can score response per impression (tratio). It can also predict Impressions, Cost Per Impression (predicted price), (phone) Response Per Impression, Web Response per impression, TRP (target rating points) and others. The list of target value types supported by the system are shown in Table x. In each case, the system uses the common set of media asset patterns defined earlier, with the ad effectiveness metric also defined earlier, to predict the target metric of interest.
For example, in order to predict Impressions, the system has expected Impressions defined for each media asset pattern type defined earlier. The system then performs a linear combination of its weighted features to predict upcoming impressions.
Exemplary Weight Training for Forecasting Impressions:
An example forecast is below for the case of impressions. Impressions don't need to undergo standardization and so the example is fairly simple. Let's say that we're trying to estimate the impressions for media instance Mi=(“Little House on The Prairie”, Hallmark, Sun 6 pm, 6/9/2013). The Media Asset Pattern Types that match this airing are shown in Table 26A and 26B below:
Exemplary Media Asset Patterns and Weights:
In one embodiment, given that there may be weights on Maptypes 86, 59, 32, and 5 with 0.25 weight each, this results in the following:
Forecast Impressions=(164,671*w1+490,169*w2+320,361*w3+232,881*w4)/sum(w1 . . . 4)=264,971
Also, assuming that the actual impressions from that airing are ultimately found to be equal to: Actual=292,497, then error can then be calculated as below:
Error=(Forecast−Actual)=27,527
Based on hundreds of thousands of examples of forecasts and actuals, the system may be trained to adjust its weights to minimize forecasting error. It may also be possible to implement variable selection process to iteratively add variables and determine if they improve the fit, and then attempt to remove variables is a similar manner to determine if there is redundance (forward-backward algorithm).
Exemplary Fatigue and Pod Adjustments During Training:
One of the objectives of the present disclosure is to accurately predict a Response Per Impression metric for a future TV broadcast. One challenge is that campaigns are rarely starting for the first time. Often the advertiser has aired their commercial on a range of different networks, and this has caused their commercial to create fatigue on these different networks.
Previous airings cause a variety of challenges for training a model to estimate future Response Per Impression. Historical data on response per impression (eg. phone response) will be distorted because of low fatigue on early airings, and high fatigue on later airings.
For example, the advertiser may have bought “Wheel of Fortune” heavily in the past. When a model is trained to predict Response Per Impression, the historical “Wheel of Fortune” will include data from when “Wheel of Fortune” was first being bought, and so the historical performance may over-estimate the performance that it may be possible to achieve if “Wheel of Fortune” is purchased today.
In order to account for fatigue, it may be desirable to adjust historical airing performance to “reverse out” the impact of fatigue. One example of how to do this is to adjust historical Response Per Impression estimates per below:
RPI_historical=RPI_historical*ln(airingcount+1)
The above fatigue adjustments should be used for ad effectiveness metrics which are related to human response, such as phone response per impression, web response per impression, and the like. Fatigue adjustments aren't needed for ad effectiveness metrics which aren't affected by human response, such as buyers per impression, or age-gender TRP estimates. These latter metrics will be the same whether or not the ad has aired in these spots previously.
Another factor which can make it difficult to predict future RPI performance is variation in historical pod position. Often media buyers negotiate rotations and may be agnostic to particular pod positioning. The pod that the ad airs in has a dramatic impact on response from the ad. The first pod has highest response, and the later the ad appears in the commercial break, the lower is the response. For the 5th ad in a commercial break, performance is just 30% of the 1st ad. This is a huge performance change, and a major variable which needs to be taken into account. One example for how to take this into consideration is to estimate RPI as a function of pod position, and then to adjust as below:
Table 27, below, shows RPI position adjustments empirically measured in a live TV campaign.
It is then possible to calculate an RPI-position1-equivalent metric by adjusting the historical RPI metrics as follows:
RPI_historical=RPI_historical(1)/RPI_historical(pod)
Exemplary RPI Adjustments During Training:
Response per impression metrics that are divided by impressions can be volatile when there are few impressions. In many cases it is possible to log-transform the RPI metric being estimated to make it robust to these outliers. This often results in far better accuracy than leaving the RPI metric un-normalized.
RPI_historical=ln(RPI_historical)
Exemplary Weights on Specific Media Asset Patterns:
A weight may be applied to an entire class of media asset patterns. For example, CNN, NBC, BRAVO, may all be weighted the same amount, and additional data encapsulated by an ad's effectiveness on CNN, NBC, and BRAVO may vary. An example of this is shown in Table 26, which describes the training process in detail. Table 26 shows an example where CNN-Tues-7 pm, CNN-Tues-8 pm, etc, all receive a weight of 0.5. The RPI score for each of these different times can of course be different, and in the example, CNN-Tues-8 pm has the highest RPM (0.5).
In one embodiment of the present disclosure, knowledge of a specific media pattern (e.g., CNN) that is equal to a value may be important for predicting an ad's effectiveness (see Table 1-3). For example, a media asset pattern of a program may be set to a weight of, e.g., 0.4. However, when the program is “The Academy Awards,” the weight may be set to 1.0. In one embodiment, special media asset patterns may be set up to cover a specific media asset pattern, and the other media asset patterns may be set to null. Table 1-3 shows an example of this: CNN-Tues-8 pm receives a weight of 0.5. This indicates that the system should “pay greater attention” to the Station-Day-Hour MAPType when the value is equal to CNN-Tues-8 pm. This is also equivalent to creating a new Media Asset Pattern Type which is equal to the specific MAP string which is being differentially weighted.
Media Asset Pattern Dummy Variable Mining:
Mining to find these special media asset patterns may involve a rule extraction algorithm. For example, the algorithm may search various search spaces, i.e., media asset patterns (station, program, genre, day, and hour). Mining may use the systems in an environment, such as the environment shown in
The system may generate every possible combination of a media asset pattern. By working from most general media asset patterns first, the system may ensure adequate “support.” Further, the system may form children media asset patterns from the general asset patterns. For example, generated media asset patterns may include: (DIY-Mon-9 pm-11 pm-Documentary); (DIY-Mon-9 pm-11 pm); (DIY-Mon); (DIY); (Documentary); (DIY-9 pm-11 pm); (Mon-9 pm-11 pm); (Mon); and (9 pm-11 pm). The system may also remove generated media asset patterns that are redundant, unlikely to be usable, and/or unlikely to be valuable, such as generated media asset patterns (Mon-9 pm-11 pm); (Mon); and (9 pm-11 pm).
There may also be constraints on a search space. Media asset patterns may be set to not allow collapsibility, which may occur if a child media asset pattern (e.g., ID-Tuesday-8 pm) is predictive, and the parent media asset pattern (e.g., ID-Tuesday) is also predictive. Thus, a child media asset pattern may be deleted (or “collapsed”), and the parent media asset pattern may be used. This may minimize a number of media asset patterns that need to be comprehended by human analysts and/or a machine learning algorithm consistent with this disclosure. This may also allow media asset patterns to work at as general a level as possible.
An example implementation may be set as follows: a media asset pattern is significant at p<0.1 level; orders from media asset pattern>=1; cost per card from media asset pattern<$10,000; and/or above average performance only.
E[Response|Media Asset Pattern]>E[Response].
An example result may be shown as shown in Table 2 below:
Table 28, below, depicts how weights can be applied to Media Asset Pattern Types as a whole, where all MAP strings receive the same weight.
Table 29, below, depicts how weights can be applied to specific Media Asset Patterns. Different MAP strings can receive different weight.
As shown in
Special Branching Structure and High-Order Features:
The model can be improved by adding structure to detect a variety of conditions. In one embodiment these conditions are implemented using a decision tree in which given a certain condition, a weighted model is executed. However these conditions could also be implemented as features themselves, incorporated as interaction terms or the like. Special conditions may include:
Table 31, below, shows trained weights for local airings, and table below that shows performance predicting local response per impression for two different advertisers.
A branch may be created, as follows:
Where <high-tratio-volatility-model> is trained on airings which are on networks that have high tratio volatility. In practice, it may be expected that the features selected for the model above will tend to have more program-specific features.
As shown in Table 32, above, TV Network FS1 has high variability in viewership for its programs even during the same day of week, hour-of-day, and program name. Variability can also be caused when networks change their schedules (eg. showing volleyball, basketball, football, etc in the same timeslots). When there is high demographic volatility as above, forecasting viewership and response from the upcoming airing will be more accurate when using program-specific features.
Table 33, below, depicts exemplary low demographic volatility networks.
Below is a sample of SQL code for calculating volatility by network
The table below shows trained model for estimating RPI for an airing which has high demographic volatility. The system makes use of Buyers per million features to increase its accuracy. Table 36, below, shows the prediction performance on airings.
It may be possible to create a special branch for syndicated airings as follows:
If <syndicated airing> then <syndicationmodel>
Tables 37A-37C below depict examples of different features used for predicting syndicated airings: Maptype 83==syndicated-station-program; maptype 83==syndicated program; maptype 76==Program-quarter of year.
Table 38, below, depicts syndicated features and degree of predictiveness for estimating response per impression where RPI is phone response per impression.
The table in
Exemplary Variable Selection:
Variable participation may be limited due to participation thresholds which remove variables, missing value handling, which enables the system to elegantly operate with missing features, and forward-backward selection, which aggressively removes variables that do not make a significant contribution to the model.
Exemplary Effects of Fatigue:
Extensive surveys and meta-studies of hundreds of publications have concluded that advertisement response shows diminishing returns when displayed to the same target audience over time. A version of the embodiment will take into account the decrease in performance during repeated exposures of advertising in the same positions, which may be referred to as estimates of “fatigue.”
One embodiment estimates fatigue as a function of individual advertisement exposures of persons participating using a panel. In this embodiment the viewers of a program are known and it may be possible to count the number of times the viewer had the TV on while the ad was on. This approach requires the existence of a panel and their viewing activity.
A second embodiment may estimate fatigue by counting airings delivered to the same program or station-time-of-day. This latter approach has an advantage in that it only requires an advertiser to keep a count of the number of airings in each media asset pattern. It does not require a panel or viewing activity in order to provide a fatigue estimate.
Another method is to use the number of historical airings in each media asset pattern and compare it to the phone response from that same media asset pattern.
Another method is to use the number of airings in media asset pattern and compare it to the web response from that media asset pattern. As the airing count increases, the web response should decrease. A Fatigue function can then be estimated and used to estimate the effect of fatigue (or of airing in the same media asset pattern).
Fatigue can also be estimated by examining set top box conversion rate versus number of exposures to an individual person. Set top box conversion rate can be calculated as the number of persons who converted (known buyers as provided by an advertiser) divided by the number of persons in the population. It may then be possible to count the converters/viewers for persons who have had 1 exposure, 2 exposures, 3 exposures and so on.
Fatigue with Airing Counts and Co-Viewing:
An airing count for media A(m) may be calculated as a count of known airings placed into media slot m. This airing count, however, may fail to take into consideration co-viewing activity. For example, an advertisement may have been run ten times on, e.g., the Military Channel's “Greatest Tank Battles.” A media buyer may wish to run the advertisement on the Military Channel's “Top 10 Aircraft,” which has had zero airings. The media buyer may have assumed such a run would avoid a decline in the advertisement's performance. However, the media buyer may be under-estimating the effective frequency.
For example, Military Channel viewers may be considered highly “insular” in their viewing habits. Thus, by airing the media buyer's advertisement in Greatest Tank Battles ten times, the media buyer may have effectively hit much of the same audience that would be viewing Top 10 Aircraft. Therefore, calculating the frequency of advertisement viewing that incorporates knowledge of co-viewing probabilities may be an important consideration.
Given knowledge of co-viewing probabilities, the probability that viewers will not have observed the advertisement may be calculated. The co-viewing probabilities may be calculated from, for example, set top box data. Thus, an effective airing rate may be represented by the following formula:
A*(mj)=max A(mi)·Pr(mi,mj)
In order to account for the impact of Fatigue, expected response per impression, rpiΩ, may be equal to the number of buyers per impression (targeting score) divided by a function of the log of airings (a number of repeat exposures), as indicated by the formula below. Thus, a targeting function may include an effect of repeat exposures.
Table 41, below, shows how Fatigue is combined with an RPI function to provide a measure of Fatigue-adjusted performance. In this case, the fatigue function is log(airingcount), and adjusted performance is RPI/log(airingcount). This can be used by media buyers to prioritize buying programs for an upcoming television campaign. This also has the effect of “intelligently” taking into account the programs where a mature TV campaign has been displayed before, and will automatically shift away from those previously purchased programs.
Table 42, as shown below, depicts cases where a target score may be calculated by combining an airing count with a targeting ratio, such as “tratio/airing count.”
Rotation Scoring:
Television media buyers often buy blocks of time on networks called “rotations.” In one embodiment of the present disclosure, these rotations are scored by the system. The rotation can be a media asset pattern instance with wildcards, or any collection of airings.
In one embodiment of the present disclosure, the system takes a “rotation” to be scored, e.g., Seattle-CNN-6 pm-9 pm, and then “explodes” this airing into each possible airing or media instance where the ad could be placed within that rotation, eg. “Seattle-CNN-6 pm-Out Front with ErinB”, “Seattle-CNN-7 pm-AC360”, “Seattle-8 pm-Piers Morgan.” These individual airings or media instances are then scored by the Scoring Service.
In one embodiment, the system assumes equal probability of the ad appearing in any of the underlying media instances.
In another embodiment, the system assumes “worst case” insertion in which it selects the underlying media instance with lowest impressions, highest CPM, lowest tratio or the like.
In another embodiment the system attempts to estimate the placement biases of the network and may distribute the airings based on the media instances with the lowest household impressions.
After scoring the underlying media instances for impressions, response per impression (tratio), buyers per impression and other scores generated by Scoring Service, the system then re-aggregates these media instances to create a final score for the rotation. In one embodiment, the system assumes equal probability and averages the underlying scores. In another embodiment, the system assumes “worst-case” insertion and so selects the media instance with the lowest impressions, highest CPM, lowest tratio or the like, and reports that back as the insertion solution for the rotation. Figure below (“Automated Media Scoring”) shows a flow-chart showing how the rotation is exploded, scored, and then each of the underlying scores put back together into a rotation score.
Table 43, below, depicts exemplary Media Asset Pattern Types matched for one airing, in which all providers are not necessarily able to carry cost, imps, etc., and where threshold drops out features if too little data exists.
Table 47, below, depicts Scoring Service Output records (examples). The records below show some examples of television airings and scored response per impression (tratio), CPM, Impressions and so on.
Table 48, below, depicts dimensions (e.g., Network ID, Program ID, Day of Week, etc.), as well as that dual feed airing may have multiple airing events (i.e., different airing dates.)
Table 50, below, depicts an example cardinality of different media asset pattern types that may be used by the system. In one embodiment there are approximately 18,642,000 pre-computed media asset patterns being used to estimate the response per impression, impressions, CPM and other aspects of a television airing.
Table 51, below, depicts an example of trained weights (wexpert) applied to each media asset pattern type. These weights are evaluated multiplied by normalized ad effectiveness scores and combined to estimate the response per impression target. Cadaline is a one-variable linear model. Cadaline_test is the model applied on a hold-out set. % is the percentage of airings where this media asset pattern type is present (non-missing). The weights below are from weightid=20.
Table 52, below, depicts yet another set of weights from an earlier model (weightid=3).
Program aspects of the technology may be thought of as “products” or “articles of manufacture” typically in the form of executable code and/or associated data that is carried on or embodied in a type of machine-readable medium. “Storage” type media include any or all of the tangible memory of the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide non-transitory storage at any time for the software programming. All or portions of the software may at times be communicated through the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer of the mobile communication network into the computer platform of a server and/or from a server to the mobile device. Thus, another type of media that may bear the software elements includes optical, electrical and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry such waves, such as wired or wireless links, optical links, or the like, also may be considered as media bearing the software. As used herein, unless restricted to non-transitory, tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates in providing instructions to a processor for execution.
While the presently disclosed sharing application, methods, devices, and systems are described with exemplary reference to mobile applications and to transmitting data, it should be appreciated that the presently disclosed embodiments may be applicable to any environment, such as a desktop or laptop computer, an automobile entertainment system, a home entertainment system, etc. Also, the presently disclosed embodiments may be applicable to any type of protocol stack.
With the above described disclosure, it may be possible to target TV ads to maximize well-defined ad response metrics at scale. As described herein, TV targeting may be defined as a well-defined supervised learning problem. Accordingly, the types of ad effectiveness methods that are available may vary, and may each be combined to offset weaknesses in each method. By combining these techniques improvements in TV ad targeting may be realized using present TV systems.
Unless the context clearly requires otherwise, throughout the description and the claims, the words “comprise,” “comprising,” and the like are to be construed in an inclusive sense, as opposed to an exclusive or exhaustive sense; that is to say, in the sense of “including, but not limited to.” As used herein, the terms “connected,” “coupled,” or any variant thereof means any connection or coupling, either direct or indirect, between two or more elements; the coupling or connection between the elements can be physical, logical, or a combination thereof. Additionally, the words “herein,” “above,” “below,” and words of similar import, when used in this application, refer to this application as a whole and not to any particular portions of this application. Where the context permits, words in the above Detailed Description using the singular or plural number may also include the plural or singular number respectively. The word “or,” in reference to a list of two or more items, covers all of the following interpretations of the word: any of the items in the list, all of the items in the list, and any combination of the items in the list.
The above detailed description of examples of the present disclosure is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed above. While specific examples for the present disclosure are described above for illustrative purposes, various equivalent modifications are possible within the scope of the present disclosure, as those skilled in the relevant art will recognize. For example, while processes or blocks are presented in a given order, alternative implementations may perform routines having steps, or employ systems having blocks, in a different order, and some processes or blocks may be deleted, moved, added, subdivided, combined, and/or modified to provide alternative or sub-combinations. Each of these processes or blocks may be implemented in a variety of different ways. Also, while processes or blocks are at times shown as being performed in series, these processes or blocks may instead be performed or implemented in parallel, or may be performed at different times. Further any specific numbers noted herein are only examples: alternative implementations may employ differing values or ranges.
The teachings of the present disclosure provided herein can be applied to other systems, not necessarily the system described above. The elements and acts of the various examples described above can be combined to provide further implementations of the present disclosure. Some alternative implementations of the present disclosure may include not only additional elements to those implementations noted above, but also may include fewer elements.
These and other changes can be made to the present disclosure in light of the above detailed description. While the above description describes certain examples of the present disclosure, and describes the best mode contemplated, no matter how detailed the above appears in text, the present disclosure can be practiced in many ways. Details of the system may vary considerably in its specific implementation, while still being encompassed by the present disclosure disclosed herein. As noted above, particular terminology used when describing certain features or aspects of the present disclosure should not be taken to imply that the terminology is being redefined herein to be restricted to any specific characteristics, features, or aspects of the present disclosure with which that terminology is associated. In general, the terms used in the following claims should not be construed to limit the present disclosure to the specific examples disclosed in the specification, unless the above detailed description section explicitly defines such terms. Accordingly, the actual scope of the present disclosure encompasses not only the disclosed examples, but also all equivalent ways of practicing or implementing the present disclosure.
This application is a continuation of U.S. patent application Ser. No. 14/586,746, filed Dec. 30, 2014, which claims the benefit of priority to U.S. Provisional Patent Application No. 61/922,007, entitled “Television Advertisement Targeting that Balances Targeting Against Previous Airings,” filed on Dec. 30, 2013, each of which is incorporated herein by reference in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 8080698 | Canning et al. | Nov 2011 | B2 |
| 8255948 | Black et al. | Aug 2012 | B1 |
| 8365213 | Orlowski | Jan 2013 | B1 |
| 20020078444 | Krewin et al. | Jun 2002 | A1 |
| 20070157228 | Bayer et al. | Jul 2007 | A1 |
| 20100100419 | Natoli et al. | Apr 2010 | A1 |
| 20110191309 | Anderson et al. | Aug 2011 | A1 |
| 20120023522 | Anderson et al. | Jan 2012 | A1 |
| 20120042338 | Kitts et al. | Feb 2012 | A1 |
| 20120042339 | Kitts et al. | Feb 2012 | A1 |
| 20120054019 | Kitts et al. | Mar 2012 | A1 |
| Number | Date | Country |
|---|---|---|
| 2013192314 | Dec 2013 | WO |
| Entry |
|---|
| Aaker, J. et al., “Nontarget markets and Viewer Distinctiveness: The Impact of Target Marketing on Advertising Attitudes,” Journal of Consumer Psychology, vol. 9, No. 3, pp. 127-140. Lawrence Erlbaum Assciates Inc., (2000). |
| Balakrishnan, S. et al., “Computational Television Advertising,” Twelfth IEEE International Conference on Data Mining, pp. 71-80, (2012). |
| Chang et al., “Consumer micro-behavior and TV viewership patterns: data analytics for the two-way set-top box,” Proceedings of the 14th Annual International ACM Conference on Electronic Commerce, pp. 272-273, ACM New York, (2012). |
| Coleman, T.F. et al., “A Reflective Newton Method for Minimizing a Quadratic Function Subject to Bounds on Some of the Variables,” Dec. 7, 1992. |
| Deshpande, R. et al., “A tale of two cities: Distinctiveness theory and advertising effectiveness,” Journal of Marketing Research, vol. 31, pp. 57-64, (1994). |
| Dietterich, T.G., “Ensemble methods in machine learning,” Oregon State University (2000). |
| Direct Response Academy, “Advanced DRTV Media: Direct Response Television Media Management, Training Course materials on Direct Television media buying,” (2008), http://www.directresponseacademy.com. |
| Jones, J., “What Does Effective Frequency Mean in 1997?”, Journal of Advertising Research, Jul., pp. 14-20, (1997). |
| Kitts, B. et al., “Targeting Television Audiences using Demographic Similarity”, 2010 IEEE International Conference on Data Mining Workshops, pp. 1391-1399, 2010. |
| Nielsen Corporation, “A2/M2 Three Screen Report: Television, Internet and Mobile Usage in the United States”, 4th Quarter 2008, (2008). |
| Nielsen Corporation, “Americans watching more TV than ever”, Nielsen Wire, May 20, 2009 http://blog.nielsen.com/nielsenwire/online—mobile/americans-watching-more-tv-than-ever/. |
| Schneider, M., “Fox wants answers from Nielsen,” http://www.variety.com/article/VR1118003924?refCatId=14, Variety, May 18, 2009. |
| Segal, A., “Nielsen Ratings: An Inaccurate Truth Out of date television ratings system exposed,” http://cornellsun.com/node/23180, The Cornell Daily Sun, Apr. 26, 2007. |
| Tellis, G. et al., “Modeling the Microeffects of Television Advertising: Which Ad Works, When, Where, for How Long, and Why?”, (2005), Marketing Science 24(3), pp. 351-366, Informs http://www-rcf.usc.edu/˜tellis/AdMicro.pdf. |
| Johansson, J.K., “Advertising and the S-Curve: A New Approach,” Journal of Marketing Research, vol. 16, No. 3, pp. 346-354, American Marketing Association (1979). |
| Simon, J.L. et al., “The Shape of the Advertising Response Function,” Journal of Advertising Research, vol. 20, No. 4, pp. 767 (2002). |
| Vakratsas, D. et al., “Advertising Response Functions Revisited: A Model of Dynamic Probabilistic Thresholds,” Marketing Science, 23(1), pp. 109-119, (2004). |
| International Search Report and Written Opinion issued in corresponding International Application No. PCT/US2014/072872, dated Jun. 11, 2015 (15 pages). |
| Number | Date | Country | |
|---|---|---|---|
| 20170223425 A1 | Aug 2017 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61922007 | Dec 2013 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 14586746 | Dec 2014 | US |
| Child | 15467411 | US |
