Universal system and method for representing and predicting human behavior

• US 8,566,256 B2
• Filed: 03/31/2009
• Issued: 10/22/2013
• Est. Priority Date: 04/01/2008
• Status: Active Grant

• Current Assignee: Certona Corporation (StartApp, Inc.)
• Original Assignee: Certona Corporation (StartApp, Inc.)
• Inventors: Hueter, Geoffrey J., Quandt, Steven C., Hueter, Noble H.
• Primary Examiner(s): RIFKIN, BEN M
• Application Number:
  US12/415,758
• Publication Number:
  US 20090248599A1
• Time in Patent Office: 1,666 Days
• Field of Search: 706/20
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06Q 10/00 Administration; Management
• CPC Class Codes: G06Q 30/00 Commerce
• CPC Class Codes: G06Q 30/0185 Product, service or business identity fraud
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: G06Q 30/0204 Market segmentation
• CPC Class Codes: G06Q 30/0255 based on user history
• CPC Class Codes: G06Q 30/0269 based on user profile or attribute
• CPC Class Codes: G06Q 30/0631 Item recommendations
• CPC Class Codes: H04L 51/212 using filtering or selective blocking

Abstract

A system and method is disclosed for profiling subjects and objects based on subjects'"'"' responses to various objects for purposes of determining and presenting the objects most likely to generate the most positive response from each visitor. Object ratings, such as aesthetic response, preference, interest, or relevancy, are explicitly submitted by subjects or derived implicitly from visitor interactions with the objects. Objects include movies, books, songs, commercial products, news articles, advertisements or any other type of content or physical item. A profiling engine processes the ratings information and generates compact profiles of each subject and object based on the similarities and differences in affinities between the group of subjects and the group of objects. A recommendation engine then generates recommendations to a subject based on similarity between the subject and object profiles. The recommendation engine can also match subjects to other subjects and objects to other objects. The recommendation engine can also predict affinity across object catalogs and across time. Additionally, the object profiles can be clustered to create behavioral object categories. The system has application in personalization, behavioral targeting, Internet retailing and interactive radio, to name but a few applications.

Claims

1. A computer-based service system comprising:

a database that contains affinity data of a plurality of users;

a computer server configured to generate a subject affinity vector representation of a subject and an object affinity vector representation of an object, and to produce an object recommendation for delivery to an external application, the produced object recommendation based on the subject affinity vector, the object affinity vector, and the affinity data;

wherein the subject affinity vector and the object affinity vector each have a respective number of dimensions; and

wherein the predicted affinity of the subject to the object is calculated by matching the subject affinity vector to the object affinity vector; and

wherein the computer server is further configured to generate the subject affinity vector and the object affinity vector by producing initial subject affinity vectors and initial object affinity vectors having respective initial dimensions, to determine predicted affinity values based on the initial subject affinity vectors, initial object affinity vectors, and a subset of the affinity data, and to calculate a cost function; and

wherein the computer server iteratively increases the dimensions of the generated subject affinity and object affinity vectors, and recalculates the cost function based on the differences between the predicted affinity values and actual affinity values, until the cost function reaches a predetermined value, and wherein the actual affinity values are based on user input.

2. The service system according to claim 1, wherein the user input is received from explicit user rating values.

3. The service system according to claim 1, wherein the user input is determined from analysis of user behavior.

4. The service system according to claim 1, wherein the database includes ratings that are produced from collected user behavior that is any of implicit behavior or explicit behavior, or both.

5. The service system according to claim 1, wherein the subject affinity vector and object affinity vector are based on data of respective subjects and objects profiled at different times and matched across time.

6. The service system according to claim 1, wherein the subject affinity vector and object affinity vector are based on data of respective subjects and objects that are clustered together in vector space, and further wherein said clusters are aligned across separate application vector spaces.

7. The service system according to claim 6, wherein the recommendations are dispersed across a targeted cluster.

8. The service system according to claim 1, wherein the number of dimensions and values of each dimension of the subject affinity vector are derived independently of and do not correspond to pre-identified attributes of the subject or specific object ratings of the subject and further wherein the number of dimensions and values of each dimension of the object affinity vector are derived independently of and do not correspond to pre-identified attributes of the object.

9. The service system according to claim 8, wherein the subject affinity vector and object affinity vector are determined from data comprising object ratings produced by the subjects, and wherein the subject affinity vectors and object affinity vectors are calculated through a number of iterations, with each iteration consisting of first calculating the subject affinity vector as the parameters of a linear regression model with the object similarity vectors as the inputs to the model and the corresponding actual affinity values as the target model outputs and then calculating the object affinity vectors as the parameters of a linear regression model with the subject similarity vectors as the inputs to the model and the corresponding actual affinity as the target model outputs, and further wherein the vector updates are iterated until either the cost function of differences between the predicted and actual affinity values or the successive decreases in such cost function are below predetermined thresholds.

10. The service system according to claim 1, wherein the external application includes any one of a web page, a personalized radio, a personal computer system, an in-store kiosk, a personal digital assistant (PDA), a mobile phone, a set top box, or a conventional telephone.

11. The service system according to claim 1, wherein the recommendation includes recommendations to a self-associated group that are calculated by combining the predicted affinities for each object across the subjects of the self-associated group.

12. The service system according to claim 1, wherein the iterative change to the generated subject vector S produced by the computer server from a time t to a time t+1 is defined, by

13. The service system according to claim 1, wherein the iterative change to the generated object vector B produced by the computer server from a time t to a time t+1 is defined by

14. The service system according to claim 1, wherein the matching the subject affinity vector to the object affinity vector consists of calculating the dot product between the subject affinity vector and the object affinity vector.

15. The service system according to claim 1, wherein combining the predicted affinities for each object across the objects of the self-associated group includes combining the predicted affinities as the geometric mean of the predicted affinities for an object across the subjects of the self-associated groups.

16. A computer-implemented method of producing an object recommendation to an external application, the method comprising;

generating, at a server computer, a subject affinity vector representation of a subject and an object affinity vector representation of an object;

producing an object recommendation for delivery from the server computer to an external application, the object recommendation based on the subject affinity vector, the object affinity vector, and affinity data of a database of a plurality of users;

wherein the subject affinity vector and the object affinity vector each have a respective number of dimensions, the predicted affinity of the subject to the object is calculated by matching the subject affinity vector to the object affinity vector, and the computer server is further configured to generate the subject affinity vector and the object affinity vector by producing initial subject affinity vectors and initial object affinity vectors having respective initial dimensions, determining predicted affinity values based on the initial subject affinity vectors, initial object affinity vectors, and a subset of the affinity data, and calculating a cost function, and wherein the computer server iteratively increases the dimensions of the generated subject affinity and object affinity vectors, and calculates the cost function based on the differences between predicted affinity values and actual affinity values, until the cost function reaches a predetermined value, wherein the actual affinity values are based on user input.

17. The method according to claim 16, wherein the user input is received from explicit user rating values.

18. The method according to claim 16, wherein the user input is determined from analysis of user behavior.

19. The method according to claim 16, wherein the database includes ratings that are produced from collected user behavior that is any of implicit behavior or explicit behavior, or both.

20. The method according to claim 16, wherein the subject affinity vector and object affinity vector are based on data of respective subjects and objects profiled at different times and matched across time.

21. The method according to claim 16, wherein the subject affinity vector and object affinity vector are based on data of respective subjects and objects that are clustered together in vector space, and further wherein said clusters are aligned across separate application vector spaces.

22. The method according to claim 21, wherein the recommendations are dispersed across a targeted cluster.

23. The method according to claim 16, wherein the number of dimensions and values of each dimension of the subject affinity vector are derived independently of and do not correspond to pre-identified attributes of the subject or specific object ratings of the subject and further wherein the number of dimensions and values of each dimension of the object affinity vector are derived independently of and do not correspond to pre-identified attributes of the object.

24. The method according to claim 23, wherein the subject affinity vector and object affinity vector are determined from data comprising object ratings produced by the subjects, and wherein the subject affinity vectors and object affinity vectors are calculated through a number of iterations, with each iteration consisting of first calculating the subject affinity vector as the parameters of a linear regression model with the object similarity vectors as the inputs to the model and the corresponding actual affinity values as the target model outputs and then calculating the object affinity vectors as the parameters of a linear regression model with the subject similarity vectors as the inputs to the model and the corresponding actual affinity values as the target model outputs, and further wherein the vector updates are iterated until either the cost function of differences between the predicted and actual affinity values or the successive decreases in such cost function are below predetermined thresholds.

25. The method according to claim 16, wherein the external application includes any one of a web page, a personalized radio, a personal computer system, an in-store kiosk, a personal digital assistant (PDA), a mobile phone, a set top box, or a conventional telephone.

26. The method according to claim 16, wherein the recommendation includes recommendations to a self-associated group that are calculated by combining the predicted affinities for each object across the subjects of the self-associated group.

27. The service system according to claim 16, wherein the matching the subject affinity vector to the object affinity vector consists of calculating the dot product between the subject affinity vector and the object affinity vector.

28. The service system according to claim 26, wherein combining the predicted affinities for each object across the objects of the self-associated group includes combining the predicted affinities as the geometric mean of the predicted affinities for an object across the subjects of the self-associated groups.

29. A computer-based service system comprising:

a database that contains affinity data of a plurality of users;

a computer server configured to generate a subject affinity vector representation of a subject and an object affinity vector representation of an object, and to produce an object recommendation for delivery to an external application, the object recommendation based on the subject affinity vector, the object affinity vector, and the affinity data;

wherein the subject affinity vector and the object affinity vector each have a respective number of dimensions, the predicted affinity of the subject to the object is calculated by matching the subject affinity vector to the object affinity vector, and the computer server is further configured to generate an initial subject affinity vector having an initial number of dimensions and to generate an initial object affinity vector having the initial number of dimensions, based on a subset of the affinity data, and is further configured to modify the initial subject affinity vector and initial object affinity vector with at least one additional dimension and produce predicted affinity values based on the modified initial subject affinity vector and modified initial object affinity vector, and to determine a cost function for the modified initial subject affinity vector and modified initial object affinity vector, based on the predicted affinity values; and

wherein the computer server is further configured to iteratively add additional dimensions to the generated subject affinity vectors and to the generated object affinity vectors to produce modified subject affinity vectors and a modified object affinity vectors and new predicted affinity values, and to determine a new cost function for the modified subject affinity vectors and the modified object affinity vectors, based on the new predicted affinity values, such that the computer server halts the modifying and uses the last modified subject affinity vector and last modified object affinity vectors as the subject affinity vector and the object affinity vector for the object recommendations, in response to a value of the cost function being equal to or less than a predetermined value of the cost function.