Patent Application
20100049586
1. Field of the Invention
Aspects of the present invention relate generally to a method for determining the best ads to present, and their positions, based on the maximization of an expected utility function.
2. Description of Related Art
Search engines (e.g., Yahoo! Search) generally show advertisements together with the results of searches, and such advertisements can be shown both inline with the results and in various other positions on the search results page (e.g., above the results, to the right of the results, etc.). Generally, when a user clicks on an ad, the advertiser pays some amount of money to the search company for the click.
Conventionally, the decision as to which ads to show is based on some metric, a score, which is generally correlated with the clickability of the ad (i.e., the probability that the user will actually click on the ad). In Internet search advertising, there are usually a limited number of ads to be displayed at any one time, in any of a number of limited positions. Display constraints can affect clickability and other metrics.
It would be desirable to predict the best ads and their corresponding presentation based on an expected utility function that can balance various factors, including, for example, revenue, cost, etc.
In light of the foregoing, it is a general object of the present invention to provide search engines with the highest value ads and related presentation, based on the maximization of an expected value of a utility function.
Detailed descriptions of one or more embodiments of the invention follow, examples of which may be graphically illustrated in the drawings. Each example and embodiment is provided by way of explanation of the invention, and is not meant as a limitation of the invention. For example, features described as part of one embodiment may be utilized with another embodiment to yield still a further embodiment. It is intended that the present invention include these and other modifications and variations.
Aspects of the present invention are described below in the context of generating the most effective set of available ads in response to a search query, including their position on the page and their relative ordering.
Throughout this disclosure, reference is made to “system,” which is used to denote an advertising/search infrastructure through which an Internet advertising network operates (e.g., Yahoo's® Publisher Network, etc.). There are currently numerous advertising infrastructures (e.g., those run by Yahoo!®, Google™, etc.) and most offer similar services, such as, for example, the serving or presenting of advertisements; “serving” or “presenting,” as used herein, is the mechanism by which advertisements are delivered to web pages. Generally, the advertising infrastructure includes or is linked to a search engine, which displays search results together with possibly relevant advertisements bought against the search.
Throughout this description, reference is made to a “query,” which is used to denote a search query given by a user when performing a search through a search engine. A query can comprise terms (or keywords), and may contain a single term, multiple terms, a phrase of terms, etc.
Throughout this description, reference also is made to a “slate,” which is used to denote a particular ad, or set of ads, in a particular position, or positions on a web page, and may also include the relative ordering of those ads. It may be the case that a slate can contain no ads (i.e., the “null” slate). The number of possible slates for any given web page (for example, a search results page) is a function of at least the number of ads available to potentially be displayed on the particular page (according to, for example, relevancy of the ad to the search query, etc.), and the number of positions available at which to show the ads, though it will be appreciated that various other constraints also may inform the possible number of slates.
Assuming, with regard to the example slates illustrated in
Referring again to
To determine a score for each possible slate, an expected utility function may be defined, wherein the expected utility function calculates an expected utility of each slate. The system may then choose the slate that maximizes the expected utility function. Though the expected utility function may use any applicable and available variables, it will generally at least attempt to balance the expected revenue against the “cost” of showing ads to the user. The cost generally represents the impact of a particular slate on the users, and may be informed by various and multiple factors. For example, if ad block 205 in
The value of an expected utility may also be a function of “unknown” variables, such as, for example, whether the user will actually click on an ad, whether there will be a post-click conversion, etc. As is known in the art, conversions are essentially user actions desired by advertisers after a user has clicked through to a site from an ad; such actions may include subsequent product sales, newsletter subscriptions, membership registrations, software downloads, etc. Obviously the probability of a click, conversion, etc., can themselves depend on several factors, including matching the right visitor with the right ad (i.e., the relevance of an ad to a particular user), the user's current interest level, the ease with which the desired action may be taken, etc.
In order to account for all of the different unknown variables, there may be a different utility function defined for each combination of unknown variables, which utility function assumes these variables are known (e.g., assuming it was known that a user clicked on a particular ad, etc.), as illustrated by Table 1 below.
As shown by the example in Table 1, a slate's utility may be determined by the actions taken with respect to that slate by a user. For example, consider the situation where a slate is presented to a user, the user actually clicks on an ad within the slate, and then the user takes some desired action at the landing page (i.e., there is a conversion); in such a case, and using the example shown in Table 1, the utility of the slate may be quantified as the clicked ad's price-per-click (PPC) value minus the cost of presenting the slate, where cost is defined by the system and based on various factors, as previously described. Similarly, where there is a click, but no conversion, the utility of the ad may be quantified as (1-D)*PPC—cost, where D is a discount factor. The discount may be applied as a fraction of the PPC, a fixed amount subtracted from the PPC, etc.
In light of these particular variables, an expected utility function may be defined, against which the various slates will be calculated. For example, an expected utility function, taking into account the information from Table 1, may take the following form:
It will be appreciated that Equation (1) is just one example of an expected utility function that takes into account the utility functions of Table 1, and that the variables of Table 1 can be removed and/or supplemented by various others and the equation shaped in any of a number of different ways so as to give any one element more weight than any other element, etc.
P in equation (1) stands for probability: whether a user will click on an ad, whether there will be a subsequent conversion, etc., cannot be known beforehand (i.e., before the ads are shown to the user), but the probability of any of those things happening can be determined to some extent (by, for example, standard machine learning or statistical methods such as logistic regression) and subsequently used to inform the expected utility function. For example, the probability that a user will click on an ad may be informed by several factors including the relevance of the ad to a particular user, where in a list of ads the particular ad is placed (i.e., its rank), whether the ad is placed above or to the side of the main content on the page, etc. Similarly, the probability that a click will turn into a conversion may turn on multiple factors. Using statistical models developed by the system over time, the probability that any of the unknown variables will result in a particular value (e.g., the probability that an ad will be clicked on) can be estimated, and this estimation can be built into the expected utility function, as is shown by equation (1).
In an embodiment, the values of the known (e.g., cost) and unknown (e.g., probability of a certain action) variables may be informed by user-specific characteristics. As an example, consider the determination of the probability that a user will click on an ad; such a determination may be informed by various user-specific characteristics, including the user's search history, (e.g., previous n queries, time stamps of the last n queries, etc.), information provided by a cookie residing on the user's system, geographic information (including, for example, the user's physical location and the current weather at that location, etc.), historical click-through rates, gender, languages spoken, marital status, number of children, income, time of day, time of year (i.e., season), etc. More specifically, and continuing with the example, the system may determine—using the user-specific characteristics available to it—that the probability that a particular user will click on an ad (or a particular ad, etc.) is greater between the hours of 9-11 PM than between the hours of 1-2 PM. The user-specific information may itself be informed by information known by the system about other users of the system who share with the particular user similar interests, backgrounds, recorded behavior, etc. It will be appreciated that the user-specific characteristics discussed above are only examples, and that various other characteristics and information will be apparent to those skilled in the art.
After the expected utility function has been defined, the system can run the function against slates corresponding to the particular query terms or the content on the web page where the ads are to be shown (e.g., where each slate to potentially be used comprises only ads that are relevant to a query, etc.), and then choose and present the slate with the maximum expected utility. It will be appreciated that various methods may be used to evaluate a subset of potential slates, rather than performing an exhaustive evaluation of every potential slate. It also will be appreciated that a slate's expected utility may be determined in a number of different ways. For example, it may be the sum of the expected utilities of each of its ads, the average or mean of the expected utilities of each of its ads, the expected utility of the entire slate (calculated after aggregate values are deduced for each of the known and unknown variables), etc.
Before an expected utility function can be defined at block 310, some threshold decisions must be made, as shown at blocks 300 and 305. First, it must be determined which unknown variables will be used to inform the utility function. After that determination, a utility function is defined for every combination of the selected variables, also taking into account the known variables (e.g., cost and revenue, as shown previously by the example in Table 1). It will be appreciated that the utility function for one combination of variables may be the same as the utility function for another combination—especially when there are numerous variables to be taken into account—and that these functions may be changed at any time depending on what utility metric the utility functions are ultimately meant to provide (e.g. revenue vs. cost, cost only, etc.). It will be appreciated that a utility value for the “null” state (i.e., when no ads are to be presented) can be defined (e.g., such a state may be ascribed zero utility); if no other state is found with a higher expected utility, then the null state would “win” and no ads would be presented to the user.
After the variables and the utility functions associated with their various combinations have been defined, an expected utility function may be defined which takes into account the various utility functions, as illustrated at block 310. The expected utility function seeks to determine the expected utility of each combination of the selected variables, based generally on the probability that various actions will be taken by the user (e.g., that the user will click on a particular ad, that a user will sign up to a service offered through a particular ad, etc.). As previously discussed, Equation (1) is an example of an expected utility function, which is based on the example variables and associated example utility functions shown in Table 1.
At block 315 a query is received from a user and the system determines a base set of advertisements that might be shown against the query; generally, the ads to be shown against the search are those that are at least somewhat relevant to the search query, the user, the user and query, etc. For example, it probably would not be useful to show an ad for diapers when the user has searched for information about cars, and so diaper ads may be filtered out of the set of ads to possibly be shown to the user (however, if such an ad was shown, it would likely be associated with a very high cost given its low relevancy). As discussed above, the relevance of particular ads to the user may be based on user-specific characteristics derived from information already known about the user, through, for example, the user's past use of the system. Once the ads have been decided, the system then determines the possible slates for those ads, as shown at block 320. The expected utility function defined at block 310 is then run against every slate to determine an expected utility of each slate, as illustrated by blocks 325 and 330. The expected utilities of each slate may be calculated in various ways.
For example, the expected utility function may be applied individually against each ad in the slate, in which case the expected utility of the slate may be the sum of the expected utilities of its constituent ads. In such a case, the costs and revenues associated with each individual ad (assuming those are the known variables used by the expected utility function) may or may not be predicated on the other ads in the slate. For example, an ad's individual cost may be a function of its position relative to the other ads (e.g., other things being equal, the last ad in a list of four ads would have a higher cost than the first ad in the list); or, cost may be based solely on the size of the ad, irrespective of where it appears on the page, or where it is with respect to the other ads in the slate.
As another example of how the expected utility of a slate may be calculated, the expected utility function may be taken across the entire slate at once, instead of calculating the expected utility of each ad in the slate and then summing them. In this case, the cost value used in the expected utility function shown in Equation (1) may be the aggregate cost associated with the screen real estate taken up by the current slate (with or without deference to other values that may inform the cost of each ad, such as, for example, relevance to the user). It will be appreciated that this value can be derived in any of a number of other different ways. Referring again to Equation (1) as an example, revenue of the slate may be determined as the average or total PPC of all the ads in the slate. It will be appreciated that this value can be derived in any of a number of other different ways. Similarly, the unknown variables may be derived. For example, going back to Table 1, the total probability that any ad will be clicked may be a function of the probabilities of each of the individual ads being clicked. Once the aggregate or total values of all the unknown and known variables used by the particular expected utility function have been determined, the expected utility function can then be run on the slate and an expected utility for the slate output.
To mitigate the cost or speed of operation or for other reasons, the system may make assumptions as to how the user may act. For example, the system may assume, based on past action, that if a user clicks on an ad in a slate, the user will not click on another ad in the same slate (e.g., the user will not hit the back button on her browser and click another ad, etc.). In such an instance, the probability that a user will click on any particular ad may become P(particular_ad)*[1-P(ad1)]*[1-P(ad2)]* . . . [1-P(adn)]. As another example, the system may assume that a user will not click on more than two ads an hour, and may adjust accordingly the cost of presenting ads to that user after two ads have been clicked on within any one-hour period (e.g., the system may remove the cost variable from the expected utility function, etc.).
No matter the decision as to how the expected utility of each slate is to be calculated, block 335 presents the slate with the highest expected utility after all of the generated slates have been processed.
The sequence and numbering of blocks depicted in
Several features and aspects of the present invention have been illustrated and described in detail with reference to particular embodiments by way of example only, and not by way of limitation. Those of skill in the art will appreciate that alternative implementations and various modifications to the disclosed embodiments are within the scope and contemplation of the present disclosure. Therefore, it is intended that the invention be considered as limited only by the scope of the appended claims.
