Granular data for behavioral targeting using predictive models

1. A computer-implemented method of targeting comprising:

receiving a plurality of granular events, wherein a granular event comprises a type that defines an on-line activity between a client and an entity;

preprocessing the received granular events to determine an amount of informational content of the granular event for target prediction, wherein the amount of informational content comprises at least one of a page view, an advertisement click, a link selection, a search query, a form completion, a posting of text, and an execution of a transaction;

generating, in a computer, preprocessed data to facilitate construction of a model based on the granular events by clustering the granular events into a number of clusters based on the informational content for target prediction, wherein said preprocessed data comprises input features;

generating a predictive model from said preprocessed data, the predictive model for determining a likelihood of a hypothetical user action, wherein the predictive model includes:

a weight for the hypothetical user action,

model parameters comprising linear combinations of said input features;

training the predictive model by tuning the weight to optimize performance of the predictive model;

selecting a user from a plurality of users;

applying the predictive model to the selected user;

scoring the user by using the predictive model; and

scoring the user by using a Poisson type model based on the ratio between a predicted number of ad clicks and an estimated number of ad views.

2. The method of claim 1 , further comprising:

ranking the user in relation to other users by using the scoring;

generating a set of ranked users; and

identifying a subset of ranked users for additional targeting steps.

3. The method of claim 1 , wherein the groups are determined by at least one of: search, search-click, sponsored search-click, page view, advertisement view, and ad-click.

4. The method of claim 1 , wherein the clustering preserves information about a predicted target, wherein the clustering comprises an automated process.

5. The method of claim 1 , wherein the preprocessing further comprises at least one of:

Pruning the received granular events, wherein pruning comprises eliminating a granular event that occurs fewer than a predetermined number of times; and

aggregating the granular events over time such that aggregating results in one total count for each granular event; and

filtering the received granular events, wherein filtering comprises removing granular events that do not meet a pre-defined criteria.

6. The method of claim 1 , further comprising: classifying the user action into a plurality of classes, the classification comprising a distribution for received granular events.

7. The method of claim 6 , the classifying comprising a binary distribution.

8. The method of claim 1 , wherein the on-line activity comprises one of the following:

viewing a web page, clicking on a link in the page,

clicking on an advertisement in the page, issuing a search query,

using a search engine, filling out a form, posting, rating a page, rating a product, and

performing a transaction.

9. The method of claim 1 , wherein the on-line activity comprises searches, the method further comprising tracking a number of clicks on a search result.

10. The method of claim 1 , wherein the on-line activity comprises page views, the method further comprising: counting, for each page in a set of pages, a number of page views.

11. The method of claim 1 , further comprising: for a predetermined period of time, predicting the number of user ad clicks and ad views.

12. The method of claim 1 , further comprising an ad group having a plurality of advertisements.

13. The method of claim 1 , wherein the predictive model comprises one or more of a support vector machine, a Bayesian type machine, a maximum entropy network, a logistic regression machine, and a linear regression model.

14. The method of claim 1 , further comprising using a Poisson type model with a parameter comprising a linear combination of granular event counts, the event counts within a behavioral history.

15. The method of claim 1 , wherein the on-line activity comprises one of: search-click, sponsored search-click and ad-click, the method further comprising: predicting for the user a propensity to click by using the predictive model.

16. The method of claim 1 , further comprising: scoring the user based on a linear combination of granular event counts from a behavioral history.

17. The method of claim 1 , further comprising: scoring the user based on a ratio between an estimated probability of being a clicker and an estimated probability of being a non-clicker.

18. The method of claim 17 , wherein the probabilities are a product of class-conditional probabilities for individual features.

19. A computer-implemented method of targeting comprising:

receiving a plurality of granular events, wherein a granular event comprises a type that defines an on-line activity between a client and an entity;

preprocessing the received granular events to determine an amount of informational content of the granular event for target prediction, wherein the amount of informational content comprises at least one of a page view, an advertisement click, a link selection, a search query, a form completion, a posting of text, and an execution of a transaction;

generating preprocessed data to facilitate construction of a model based on the granular events by clustering the granular events into a number of clusters based on the informational content for target prediction, wherein said preprocessed data comprises input features;

generating, in a computer, a predictive model from said preprocessed data, the predictive model for determining a likelihood of a hypothetical user action, wherein the predictive model includes:

a weight for the hypothetical user action,

model parameters comprising linear combinations of said input features;

training the predictive model by tuning the weight to optimize performance of the predictive model;

selecting a user from a plurality of users;

applying the predictive model to the selected user;

scoring the user by using the predictive model; and

scoring the user by using a Poisson type model with a parameter comprising a linear combination of granular event counts, the event counts within a behavioral history.

20. A system for targeting comprising:

a computer apparatus comprising a hard drive, processor, memory, and an execution module for executing instructions comprising the steps of:

receiving a plurality of granular events, wherein a granular event comprises a type that defines an on-line activity between a client and an entity;

preprocessing the received granular events to determine an amount of informational content of the granular event for target prediction, wherein the amount of informational content comprises at least one of a page view, an advertisement click, a link selection, a search query, a form completion, a posting of text, and an execution of a transaction;

generating preprocessed data to facilitate construction of a model based on the granular events by clustering the granular events into a number of clusters based on the informational content for target prediction, wherein said preprocessed data comprises input features;

generating a predictive model from said preprocessed data, the predictive model for determining a likelihood of a hypothetical user action, wherein the predictive model includes:

a weight for the hypothetical user action,

model parameters comprising linear combinations of said input features;

training the predictive model by tuning the weight to optimize performance of the predictive model;

selecting a user from a plurality of users;

applying the predictive model to the selected user;

scoring the user by using the predictive model; and

scoring the user by using a Poisson type model based on the ratio between a predicted number of ad clicks and an estimated number of ad views.