Interactive hybrid recommender system

• US 8,019,707 B2
• Filed: 09/20/2007
• Issued: 09/13/2011
• Est. Priority Date: 09/20/2007
• Status: Expired due to Fees

• Current Assignee: Deutsche Telekom AG
• Original Assignee: Deutsche Telekom AG
• Inventors: Piratla, Nischal, Shani, Gai, Meisels, Amnon, Rokach, Lior
• Primary Examiner(s): Gaffin; Jeffrey A
• Assistant Examiner(s): GONZALES, VINCENT
• Application Number:
  US11/903,094
• Publication Number:
  US 20090164442A1
• Time in Patent Office: 1,454 Days
• Field of Search: 706/46, 706/45
• US Class Current: 706/45
• CPC Class Codes: G06F 16/335 Filtering based on additional data, e.g. user or group profiles filtering in web context G06F16/9535, G06F16/9536

Abstract

A hybrid recommender system, in which the initial stereotype is manually defined by an expert and an affinity vector of stereotypes relating to each specific user who registers onto the system, is created to define a specific profile for each user. Recommendations for a specific user are generated according to the initial stereotype and the affinity vector of stereotypes. A binary feedback, from user regarding specific items picked by him is received (e.g., while of the item), which can be either positive or negative. Then the affinity vector of stereotypes is updated.

Claims

1. A hybrid recommender system for delivering content that fits a user preferences, implementing a process comprising the steps of:

a. manually defining one or more initial stereotypes by an expert;

b. creating an affinity vector of stereotypes relating to each specific user who registers onto the system, thereby creating a specific profile for each user, wherein said affinity vector for each user is created according to her response to a set of questions, where each question compares several representative items, and a form, based on a decision tree and specifying said user'"'"'s interests, and wherein said affinity vector is saved in a database;

c. generating recommendations for a specific user according to said initial stereotype and said affinity vector of stereotypes, wherein said recommendations are generated using a matching engine that computes the relevance of an item to the user, thereby creating a recommendation list of items to incorporate into the database;

d. receiving feedback from user regarding specific items picked by him, wherein said feedback can be either positive or negative; and

e. updating said affinity vector of stereotypes and said recommendation list of items saving them in said database;

wherein the content delivered to the user contains a predefined number of the most relevant items to her according to said recommendation list.

2. The hybrid recommender system according to claim 1 wherein the user answers between none and all of the questions appearing on the form.

3. The hybrid recommender system according to claim 1 wherein the form is partially filled out according to information available to the system from sources other than the user.

4. The hybrid recommender system according to claim 1 wherein said form is based on a decision tree.

5. The hybrid recommender system according to claim 1 wherein the form is learned by a tree induction process thereby creating a new decision tree for each new user.

6. The hybrid recommender system according to claim 1 wherein the recommendations for a specific user are generated using a matching engine that computes the relevance of each item (i) to said user (u) with affinity vector (v) according to

7. The hybrid recommender system according to claim 1 wherein the recommendation comprises items with unexpected context to the user chosen according to an ε

-greedy technique and/or by randomly boosting the relevance of a random stereotype.

8. The hybrid recommender system according to claim 1 wherein the feedback is binary.

9. The hybrid recommender system according to claim 1 wherein the feedback uses explicit ratings, implicit ratings or both implicit and explicit ratings.

10. The hybrid recommender system according to claim 1 wherein the feedback is received from the user during her use of the item.

11. The hybrid recommender system according to claim 1 wherein the feedback is automatically registered by said system according to the user'"'"'s reaction to the recommended items.

12. The hybrid recommender system according to claim 1 wherein the affinity vectors are updated by a computationally light process, a computationally heavy process, or both said processes.

13. The hybrid recommender system according to claim 12 wherein the affinity vectors are updated by a computationally light process comprising the steps of:

a. updating the affinity vector according to feedback received from the user;

b. re-computing the recommendation list of item to incorporate new items added into the database, wherein a new relevance value for each new item (i) is computed to the user (u) with affinity vector (v) according to

14. The hybrid recommender system according to claim 13 wherein the number of clusters is either predefined or automatically induced.

15. The hybrid recommender system according to claim 13 wherein the clustering process is a fuzzy C-Means clustering process.

16. The hybrid recommender system according to claim 12 wherein the computationally heavy updating process comprises the steps of:

a. automatically constructing a user representation profile that comprises a list of preferred values for each possible attribute by observing the list of items rated positively;

b. creating clusters of similar users by way of a clustering process using a distance metric over the similarity of said user representation profile;

c. creating a stereotype profile wherein the users are weighted according to said distance metric; and

d. reconstructing an affinity vector for each user according to said stereotype profile.

17. The hybrid recommender system according to claim 16 wherein a soft-clustering process is used.

18. The hybrid recommender system according to claim 1 wherein the manually defined initial stereotypes are maintained and used to define new users.