Patent Application
20090271228
An incredible amount of information is accessible to individuals who have access to a networked device. Pursuant to an example, a user can search for a particular topic by proffering a search query to a search engine. The search engine, utilizing the proffered query, can locate and rank numerous web pages and provide such pages to the user. Therefore, for instance, a web page deemed most relevant to the user (given the proffered query) will be displayed most prominently to the user, while other less relevant pages will be displayed less prominently.
Along with facilitating location of information, the Internet is being used for generation of revenue. For instance, a retailer can create a website that is designed for the sale of goods and services offered by the retailer. In addition, websites exist that are dedicated to auctioning goods and/or services offered by retailers and/or individuals. Oftentimes, consumers prefer purchasing items online, as they can avoid hassles associated with driving to shopping centers.
Another manner in which the Internet has been used to generate revenue is through sale of advertisements that are displayed on web pages. For instance, when a user proffers a query to a search engine, contents of the query can be made available to prospective advertisers. The advertisers purchase space on a web page that shows search results based at least in part upon contents of the query. For instance, if the user searches for “digital camera”, a retailer that sells digital cameras may wish to provide an advertisement to the user in hopes that the user will purchase a digital camera from the retailer. Revenue can be generated by the search engine, for instance, if the user selects the advertisement. Web pages can also sell space to advertisers to generate revenue for the owner of the web page. Conventionally, online advertising relies on immediate context for selecting relevant advertisements to display to users. Immediate context may include a current search query, queries proffered by a user in a single session, and page content of a web page where an advertisement is displayed.
The following is a brief summary of subject matter that is described in greater detail herein. This summary is not intended to be limiting as to the scope of the claims.
Various technologies pertaining to targeted advertising are described herein. More particularly, generation of a user profile (that is customized for a user based upon previous behaviour) for utilization in connection with targeting advertisements to a user is described herein. The user profile may include numerous keywords, wherein each of the keywords may optionally be assigned a score or a plurality of scores. In an example, a score assigned to a keyword may be indicative of a probability that an advertisement corresponding to the keyword will be monetized by the user. For instance, the probability may relate to a probability that the user will select an advertisement that is served to the user, wherein an advertiser serves the advertisement based at least in part upon a keyword in the user profile.
As noted above, the user profile may be generated based upon previous user behaviour, such as behaviour that can be ascertained from search engine logs, queries proffered by a user, web pages visited by the user, advertisements selected by the user, amongst other data. A first set of keywords (e.g., a word or a collection of words) can be extracted from the user behaviour data. Thereafter, keywords that are related to the first set of keywords can be automatically ascertained. Pursuant to an example, machine-learning techniques can be utilized to determine the related keywords. The first set of keywords and the related keywords in combination can be referred to as a raw user profile.
A user profile that includes a plurality of keywords may then be generated based at least in part upon contents of the raw user profile. Pursuant to an example, machine learning techniques can be utilized in connection with generating the user profile, which includes selecting keywords to include in the user profile as well as individually assigning scores to keywords in the user profile. Once the user profile is created, advertisements can be served to a user based at least in part upon content of the user profile.
Other aspects will be appreciated upon reading and understanding the attached figures and description.
Various technologies pertaining to generating a user profile that can be employed in connection with targeted advertising will now be described with reference to the drawings, where like reference numerals represent like elements throughout. In addition, several functional block diagrams of example systems are illustrated and described herein for purposes of explanation; however, it is to be understood that functionality that is described as being carried out by certain system components may be performed by multiple components. Similarly, for instance, a component may be configured to perform functionality that is described as being carried out by multiple components.
With reference to
In still yet another example, the user data 104 may include off-line activity, such as location information received from position sensors, credit card purchase information, television viewing habits of the user, and/or other information. Additionally, the user data 104 may include online purchases of the user (e.g., identification of purchased items and/or services), time of day that a query was proffered to a search engine, data pertaining to music reviewed and/or listened to by the user, contextual data corresponding to online activity, etc.
A profile generator component 106 can receive a subset of data from the user data 104 and can generate a user profile 108 based at least in part upon the subset of data from the user data 104. The user profile 108 includes a plurality of keywords, wherein each of the keywords may individually be assigned a score or scores. A keyword may be a single word or a plurality of words and/or characters in a certain sequence. In an example, a score that is assigned to a keyword can be indicative of a probability that an advertisement corresponding to the keyboard will be monetized. As used herein, the term “monetized” may mean realization of revenue by any suitable manner, such as receiving payment upon a user clicking upon an advertisement, receiving payment for a keyword, and/or the like.
Pursuant to an example, a score assigned to a keyword in the user profile 108 can indicate a probability that the user will select an advertisement that is served by an advertiser in response to the keyword. In another example, a score assigned to a keyword in the user profile 108 may indicate revenue expected to be received from an advertiser who will place a bid on advertising space when the keyword is received. In yet another example, a score assigned to a keyword in the user profile 108 may be indicative of a probability that the user will be interested in topics corresponding to the keyword at some point in the future. The profile generator component 106 can be “tuned”, for instance, to assign scores based upon any one or combination of the aforementioned examples. Furthermore, the profile generator component 106 can infer keywords to include in the user profile 108 that are expected to be of interest to a user at a future point in time, wherein the profile generator component 106 infers the keywords based at least in part upon the user data 104. Generation of the user profile 108 will be described in greater detail below.
In a particular example, the user profile 108 may be limited to a threshold number of keywords. For instance, raw data logs may include several thousand keywords for any given user, and it may be computationally burdensome to search through advertisements submitted for each keyword in connection with advertisers purchasing advertising space on a graphical user interface. Thus, the user profile 108 may be compact and include a relatively small number of keywords. For instance, the user profile 108 may include five keywords. In another example, the user profile 108 may include twenty keywords. In yet another example, the user profile 108 may include five hundred keywords. A number of keywords in the user profile 108 may be dependent upon computational considerations, and may vary over time.
Referring now to
A related keyword determiner component 206 can receive keywords determined by the keyword determiner component 202 and can determine keywords that are related to the received keywords (related keywords). These related keywords may, for instance, better reflect a user's future interests than keywords from the user data 104. In an example, a keyword “tree” may be most suitable for a user whose search queries include “oak” and “pine”, even though the keyword “tree” was never proffered by the user as a query. The keyword determiner component 206 may use any suitable technique or combination of techniques to determine related keywords.
The receiver component 102 can receive keywords from the keyword determiner component 202 and/or related keywords from the related keyword determiner component 206. The combination of keywords determined from the user data 104 and keywords related thereto (as determined by the related keyword determiner component 206) can be referred to as a raw user profile 208. The profile generator component 106 receives the raw user profile 208 and generates the user profile 108 based at least in part upon contents of the raw user profile 208.
More specifically, the profile generator component 106 can generate the user profile 108 such that it includes a non-empty set of keywords, wherein each of the keywords may be individually assigned a score (or scores) that is indicative of, for instance, a measure of relevance of the keyword to the user. As noted above, a score assigned to a keyword may be indicative of a probability that an advertisement served in response to the keyword will be monetized (e.g., selected) by a user corresponding to the user profile 108. In another example, a score assigned to a keyword may be indicative of a probability that the user will be interested in a topic represented by the keyword at some point in the future.
The profile generator component 106 can consider a multitude of factors when generating the user profile 108 based at least in part upon keywords in the raw user profile 208. For instance, the profile generator component 106 can select a keyword to include in the user profile 108 based at least in part upon recency and/or frequency of occurrence of the keyword in the user data 104. In another example, the profile generator component 106 can select a keyword to include in the user profile 108 based at least in part upon monetization performance of advertisers/advertisements that bid on the keyword. In another example, the profile generator component 106 can select a keyword to include in the user profile 108 based at least in part upon demographic performance of the keyword. Other factors that may be considered by the profile generator component 106 include a representativeness of a keyword with respect to other keywords in the raw user profile, a likelihood that a keyword reflects user interests that will remain relevant in the future given past keyword occurrences, a propensity of a keyword to generate advertisements that the user is likely to select, amongst other factors. The profile generator component 106 may take into consideration one or more of the aforementioned factors (in any suitable combination) when generating the user profile 108.
The profile generator component 106 may quantify a factor by analyzing features of keywords. In an example, the profile generator component 106 may compute the following feature for each keyword in the raw user profile 208 to capture the frequency and recency of occurrences of the keyword in the user data 104, as well as those of related keywords:
where t(oi) represents a function that quantifies recency for every occurrence oi of a given keyword kw, and w(kw, kw′) represents a similarity weight between the keyword kw and its neighbor kw′. An example function that penalizes recency is as follows:
t(oi)=1−α log(NumberOfDaysSinceOccurrence(oi)).
where α<1 is a constant.
Of course, the profile generator component 106 can compute other features that are constructed to quantify the aforementioned factors as well as other properties of keywords (or a user) that may be of interest, such as overall advertisement click-through probability.
As keywords can be associated with numerous features, it may be difficult to determine relative importance of features or manually construct a function that combines features into a single score that can be assigned to a keyword. Accordingly, the profile generator component 106 can use a machine-learned function/algorithm to select/generate keywords to be included in the user profile 108. Example machine learning techniques for generation of functions/algorithms that may be utilized by the profile generator component 106 are described in greater detail below.
Now turning to
With reference now to
The profile generator component 106 receives a subset of the user data 104 and a subset of the advertising information and generates the user profile 108 based at least in part thereon. As described above, the user profile 108 includes a plurality of keywords that may be individually assigned scores, wherein a score assigned to a keyword can be indicative of a probability that an advertisement corresponding to the keyword will be monetized (e.g., selected by the user and resulting in generated revenue).
The system 400 additionally includes an advertising component 404 that analyzes the user profile 108 and serves an advertisement 406 to the user based at least in part upon content of the user profile 108. Pursuant to an example, a user that corresponds to the user profile 108 may initiate a search session (e.g., by directing a browser to a search engine). The advertising component 404 can analyze the user profile 108 and serve an advertisement to the user based at least in part upon contents of the user profile 108. For instance, the advertising component 404 can review the user profile 108 and determine that the user is most likely to select an advertisement that corresponds to a keyword (e.g., a keyword in the user profile 108 that has been assigned the highest score). The advertising component 108 can serve an advertisement to the user that corresponds to the keyword prior to the user proffering a query to the search engine, or as a contextual (non-search) advertisement on any webpage.
In a detailed example, the user profile 108 may include a keyword “blue jeans”, and such keyword may be assigned a relatively high score. Accordingly, the user may be interested in clothing products in general, and blue jeans in particular. A retailer that sells blue jeans may wish to advertise to the user due to interest of the user in blue jeans—accordingly, the advertising component 404 can serve an advertisement pertaining to blue jeans to the user.
Now turning to
The system 500 further includes an updater component 504 that receives recent online and/or offline activity of the user, wherein the recent online and/or offline activity may be activity undertaken by the user after the user profile 108 was generated or last updated. For instance, recent activity of a user may indicate that interests of the user have altered, and therefore it may be desirable to alter the contents of the user profile 108 to reflect the change in interests of the user. The recent activity of the user may be received by the profile generator component 106, and the profile generator component 106 can access the user profile 108 from the data store 502. The profile generator component 106 can analyze contents of the user profile 108 and update such contents based at least in part upon the recent online activity of the user. For instance, the profile generator component 106 can add or remove keywords from the user profile 108. In another example, the profile generator component 106 can modify scores assigned to keywords based at least in part upon the recent activity.
Now referring to
The learner component 602 can also receive subsequently selected keywords 606. More specifically, the candidate keywords 604 may relate to user activity in a data log that occurred prior to a particular time Ts. The subsequently selected keywords 606 may relate to activity of the user in the data log that occurred after the time Ts. Accordingly, the subsequently selected keywords 606 are indicative of future interests of the user after time Ts.
More specifically, the learner component 602 can use supervised machine learning methods to learn a function that combines computed features of keywords into a single score that, for instance, is indicative of a probability that an advertisement corresponding to a keyword will be monetized. Supervised machine learning methods can learn (approximate) a function S that maps input in some form (e.g., a set of keywords, where each keyword is represented by feature values) to an output (e.g., a most desirable set of keywords that represent future user interests that correspond to optimal advertising revenue).
The learning component 602 can rely on training data to learn the function S (e.g., the candidate keywords 604 and subsequent keyword selections 606). In the example system 600, as indicated above, the training data can be obtained from past user behavior that is separated based upon some intermediate time into “observations” that include behavior before the time and “best predictions” computed on behavior after the time, wherein “best” may mean, for example, most monetizable keywords. The learner component 602 can learn the function S that produces a set closest to “best predictions” from “observations.”
The learner component 602 may use any suitable learning mechanisms in connection with learning the function S. For instance, the learner component 602 may use maximum-margin methods, probabilistic models, regression trees, and/or the like. Furthermore, the learner component 602 can output a function based upon a variety of learning objectives. For instance, the system 600 may include a data store 608 that retains advertising information 610, which may include clicks on advertisements by the user and/or several users. The learner component 602 may be constructed to produce a function S that prefers keywords likely to be clicked in the future and produce maximum revenue. In another example, the learner component 602 may be configured to output a function S that generates predictions that correspond to the most salient interests of a user. In addition, the learner component 602 can combine multiple objectives.
The learner component 602 may include or utilize different learning methods depending on a type of scoring function that is desirably learned. Some scoring functions operate on multiple keywords to produce an entire set of keywords at once. To learn such functions, the learner component 602 can employ structured learning algorithms. In a different example, a scoring function can operate on individual keywords assigned thereto and output real-valued scores. This scoring function can be learned using standard learners that predict a single value based on keyword features.
With reference now to
Moreover, the acts described herein may be computer-executable instructions that can be implemented by one or more processors and/or stored on a computer-readable medium or media. The computer-executable instructions may include a routine, a sub-routine, programs, a thread of execution, and/or the like. Still further, results of acts of the methodologies may be stored in a computer-readable medium, displayed on a display device, and/or the like. Further, it is to be understood that at least some of the acts may be supplemented by functionality, acts, and/or features of the systems/components described above.
Referring specifically to
At 706, a candidate set of keywords is generated based at least in part upon the received historical online activity of the user. As described above, the candidate set of keywords may include keywords in a log of searching and browsing activity of a user, as well as keywords related to such keywords.
At 708, a user profile is generated based at least in part upon the candidate set of keywords. The user profile may include a plurality of keywords that are each individually assigned one or more scores. For example, a score assigned to a keyword can be indicative of a probability that the keyword represents future interests of the user. In another example, a score assigned to a keyword can be indicative of a probability that an advertisement corresponding to the keyword will be monetized.
Pursuant to an example, content of a user profile, including scores assigned to keywords, may be dependent on the current context of a user. For instance, if the user is determined to be at work the user profile may include a first set of keywords, while if the user is determined to be at home the user profile may include a second set of keywords. In another example, scores assigned to keywords may alter depending on context corresponding to the user. For instance, scores assigned to keywords in the profile may depend upon user location, time of day, amount of time between a current search session and a previous search session, weather conditions, and/or the like. The methodology 700 completes at 710.
Turning now to
At 806, a first set of keywords is determined based at least in part upon the historical searching and browsing activity of the user. At 808, a second set of keywords is determined, wherein the keywords in the second set of keywords are related to keywords in the first set of keywords.
At 810, the first set of keywords and the second set of keywords are combined to create a raw user profile. In other words, the raw user profile includes the first set of keywords and the second set of keywords. At 812, advertiser data that is indicative of advertiser performance with respect to keywords in the raw user profile is received. For instance, data relating to which keywords are bid upon by advertisers, how often advertisements corresponding to certain keywords are monetized, and other advertiser data may be received at 812.
At 814, a plurality of keywords that are germane to predicted future interests of the user are determined based at least in part upon the raw user profile. The plurality of keywords determined at 814 may include keywords that are in the raw user profile as well as keywords that are not in the raw user profile. In another example, the plurality of keywords determined at 814 may only include keywords that are in the raw user profile.
At 816, scores are assigned to each of the plurality of keywords determined at 814, wherein the scores are based at least in part upon the received advertiser data. A score assigned to a keyword can be indicative of a probability that an advertisement corresponding to the keyword will be monetized when an advertisement is served to the user. At 818, a user profile is generated, wherein the user profile includes a subset of the keywords determined at 814 and scores assigned thereto. Pursuant to an example, the generated user profile can include a threshold number of keywords. For instance, the generated user profile may be limited to ten keywords. In another example, the generated user profile may be limited to fifteen keywords. In yet another example, the generated user profile may be limited to a hundred keywords. The threshold may also be set based on scores assigned to keywords. For example, the generated user profile may be limited to keywords with a score above or below a certain value or within a certain range of values. The methodology 800 completes at 820.
With reference now to
With reference now to
Now referring to
The computing device 1200 additionally includes a data store 1208 that is accessible by the processor 1202 by way of the system bus 1206. The data store 1208 may include executable instructions, user profiles, raw user profiles, keywords, scores, etc. The computing device 1200 also includes an input interface 1210 that allows users or external devices to communicate with the computing device 1200. For instance, the input interface 1210 may be used to receive instructions from an external computer device, queries from a user, etc. The computing device 1200 also includes an output interface 1212 that interfaces the computing device 1200 with one or more external devices or allows information to be provided to a user. For example, the computing device 1200 may display images, search results, advertisements, or the like by way of the output interface 1212.
Additionally, while illustrated as a single system, it is to be understood that the computing device 1200 may be a distributed system. Thus, for instance, several devices may be in communication by way of a network connection and may collectively perform tasks described as being performed by the computing device 1400.
As used herein, the terms “component” and “system” are intended to encompass hardware, software, or a combination of hardware and software. Thus, for example, a system or component may be a process, a process executing on a processor, or a processor. Additionally, a component or system may be localized on a single device or distributed across several devices.
It is noted that several examples have been provided for purposes of explanation. These examples are not to be construed as limiting the hereto-appended claims. Additionally, it may be recognized that the examples provided herein may be permutated while still falling under the scope of the claims.
