Finding predictive cross-category search queries for behavioral targeting
A method and apparatus for finding predictive cross-category search queries for behavioral targeting in a networked online display advertising system. The methods include aggregating a training model dataset, the training model dataset comprising a history of clicks corresponding to historical advertisements. The training model dataset also contains plurality of targeting categories related to the history of clicks. Various techniques are disclosed for selecting a plurality of features from the training model dataset and calculating a click probability for a subject advertisement to be clicked by a user from a page, the calculating operations using features of the page that is to be presented to the user. Embodiments include mapping a particular query to one of the targeting categories and then presenting the subject advertisement selected on the basis of the value of the click probability. Normalization scales down the value of the click probabilities to filter out false positive categories.
1. A method for finding predictive cross-category search queries for behavioral targeting, comprising:
aggregating, using a computer, at least one training model dataset formed by a particular configuration of a data structure, the training model dataset comprising multiple configured data structures each representing an advertisement impression and including at least a history of clicks corresponding to historical advertisement information, a plurality of page features including a position of an advertisement within the page as shown to a particular user, and a plurality of internet property features, and the training model dataset comprising a plurality of targeting categories derived from the historical advertisement information;
training a baseline training model dataset with an initial feature set including page information features and advertisement information features, wherein the initial feature set is used to model a prior distribution of clicks and absence of clicks in a training set;
determining historical query and targeting category pairs such that the user historical query of the pair is predictive of clicks on display ads with the targeting category of the pair;
selecting, using a computer, a plurality of features from the at least one training model dataset, wherein the selected plurality of features include initial features and at least one candidate feature, wherein the candidate feature varies to fit training data and provides measuring likelihood gain of the candidate feature when added to the baseline training model dataset;
calculating a click probability for a subject advertisement to be clicked by a user from a page, said calculating using at least the selected plurality of features, wherein the initial features include features of the page, and wherein the at least one candidate feature is different from the initial features of the at least one training model dataset, and said calculating being normalized for queries that have a high click propensity and no relation to any user interest in a behavioral targeting taxonomy; and
serving the subject advertisement to the user, when the click probability of the subject advertisement is predictive of clicks on display ads based on the determined historical query and targeting category pairs.
2. The method of claim 1 , further comprising:
mapping a particular query to at least one targeting category; and
presenting, on a computer display, the subject advertisement on the page, the subject advertisement selected on the basis of the value of the click probability.
3. The method of claim 2 , wherein the mapping is performed using only an association between a query and at least one of a history of clicks.
4. The method of claim 1 , wherein aggregating the training model dataset includes aggregating at least one of, a plurality of advertisement features, a plurality of user interest features, a plurality of internet property features, a plurality of page features.
5. The method of claim 1 , wherein aggregating the training model dataset includes aggregating a data structure including at least one of, a user cookie, a timestamp, a targeting category, a position, a property.
6. The method of claim 1 , wherein the mapping includes a normalization operation.
7. The method of claim 1 , wherein the selecting includes at least one of, a threshold feature, a top n feature, a CTR ratio feature, a top n gain feature, an in-category feature.
8. The method of claim 1 , wherein the selecting is performed using a click prediction accuracy evaluator.
9. An advertising server network for finding predictive cross-category search queries for behavioral targeting, comprising:
a module for aggregating, using a computer, at least one training model dataset formed by a particular configuration of a data structure, the training model dataset comprising multiple configured data structures each representing an advertisement impression and including at least a history of clicks corresponding to historical advertisement information, a plurality of page features including a position of an advertisement within the page as shown to a particular user, and a plurality of internet property features, and the training model dataset comprising a plurality of targeting categories derived from the historical advertisement information;
a module for training a baseline training model dataset with an initial feature set including page information features and advertisement information features, wherein the initial feature set is used to model a prior distribution of clicks and absence of clicks in a training set;
a module for determining historical query and targeting category pairs such that the user historical query of the pair is predictive of clicks on display ads with the targeting category of the pair;
a module for selecting, using a computer, a plurality of features from the at least one training model dataset, wherein the selected plurality of features include initial features and at least one candidate feature, wherein the candidate feature varies to fit training data and provides measuring likelihood gain of the candidate feature when added to the baseline training model dataset;
a module for calculating a click probability for a subject advertisement to be clicked by a user from a page, said calculating using at least the selected plurality of features, wherein the initial features include features of the page, and wherein the at least one candidate feature is different from the initial features of the at least one training model dataset, and said calculating being normalized for queries that have a high click propensity and no relation to any user interest in a behavioral targeting taxonomy; and
serving the subject advertisement to the user, when the click probability of the subject advertisement is predictive of clicks on display ads based on the determined historical query and targeting category pairs.
10. The advertising server network of claim 9 , further comprising:
mapping a particular query to at least one targeting category; and
presenting, on a computer display, the subject advertisement on the page, the subject advertisement selected on the basis of the value of the click probability.
11. The advertising server network of claim 10 , wherein the mapping is performed using only an association between a query and at least one of a history of clicks.
12. The advertising server network of claim 9 , wherein aggregating the training model dataset includes aggregating at least one of, a plurality of advertisement features, a plurality of user interest features, a plurality of internet property features, a plurality of page features.
13. The advertising server network of claim 9 , wherein aggregating the training model dataset includes aggregating a data structure including at least one of, a user cookie, a timestamp, a targeting category, a position, a property.
14. The advertising server network of claim 9 , wherein the mapping includes a normalization operation.
15. The advertising server network of claim 9 , wherein the selecting includes at least one of, a threshold feature, a top n feature, a CTR ratio feature, a top n gain feature, an in-category feature.
16. A non-transitory computer readable medium comprising a set of instructions which, when executed by a computer, cause the computer to find predictive cross-category search queries for behavioral targeting, the set of instructions for:
aggregating, using a computer, at least one training model dataset formed by a particular configuration of a data structure, the training model dataset comprising multiple configured data structures each representing an advertisement impression and including at least a history of clicks corresponding to historical advertisement information, a plurality of page features including a position of an advertisement within the page as shown to a particular user, and a plurality of internet property features, and the training model dataset comprising a plurality of targeting categories derived from the historical advertisement information;
training a baseline training model dataset with an initial feature set including page information features and advertisement information features, wherein the initial feature set is used to model a prior distribution of clicks and absence of clicks in a training set;
determining historical query and targeting category pairs such that the user historical query of the pair is predictive of clicks on display ads with the targeting category of the pair;
selecting, using a computer, a plurality of features from the at least one training model dataset, wherein the selected plurality of features include initial features and at least one candidate feature, wherein the candidate feature varies to fit training data and provides measuring likelihood gain of the candidate feature when added to the baseline training model dataset;
calculating a click probability for a subject advertisement to be clicked by a user from a page, said calculating using at least the selected plurality of features, wherein the initial features include features of the page, and wherein the at least one candidate feature is different from the initial features of the at least one training model dataset, and said calculating being normalized for queries that have a high click propensity and no relation to any user interest in a behavioral targeting taxonomy; and
serving the subject advertisement to the user, when the click probability of the subject advertisement is predictive of clicks on display ads based on the determined historical query and targeting category pairs.
17. The computer readable medium of claim 16 , further comprising:
mapping a particular query to at least one targeting category; and
presenting, on a computer display, the subject advertisement on the page, the subject advertisement selected on the basis of the value of the click probability.
18. The computer readable medium of claim 17 , wherein the mapping is performed using only an association between a query and at least one of a history of clicks.
19. The computer readable medium of claim 16 , wherein aggregating the training model dataset includes aggregating at least one of, a plurality of advertisement features, a plurality of user interest features, a plurality of internet property features, a plurality of page features.
20. The computer readable medium of claim 16 , wherein aggregating the training model dataset includes aggregating a data structure including at least one of, a user cookie, a timestamp, a targeting category, a position, a property.