ROBUST BAYESIAN MATRIX FACTORIZATION AND RECOMMENDER SYSTEMS USING SAME

• US 20130226839A1
• Filed: 02/27/2012
• Published: 08/29/2013
• Est. Priority Date: 02/27/2012
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Archambeau, Cedric, Bouchard, Guillaume, Lakshminarayanan, Balaji
• Granted Patent: US 8,880,439 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

In a recommender method, Bayesian Matrix Factorization (BMF) is performed on a matrix having user and item dimensions and matrix elements containing user ratings for items made by users in order to train a probabilistic collaborative filtering model. A recommendation is generated for a user using the probabilistic collaborative filtering model. The recommendation may comprise a predicted item rating, or an identification of one or more recommended items. The recommender method is suitably performed by an electronic data processing device. The BMF may employ non-Gaussian priors, such as Student-t priors. The BMF may additionally or alternatively employ a heteroscedastic noise model comprising priors that include (1) a row dependent variance component that depends upon the matrix row and (2) a column dependent variance component that depends upon the matrix column.

Claims

1. An apparatus comprising:

an electronic data processing device configured to perform a recommender method including;

performing Bayesian Matrix Factorization (BMF) on a matrix having user and item dimensions using non-Gaussian priors to train a probabilistic collaborative littering model; and

generating a recommendation for a user comprising a predicted item rating or identification of one or more recommended items using the probabilistic collaborative flittering model.

2. The apparatus of claim 1, wherein the non-Gaussian priors comprise Gaussian scale mixture priors.

3. The apparatus of claim 2, wherein the Gaussian scale mixture priors comprise Student-t priors.

4. The apparatus of claim 2, wherein the BMF is performed using variance values for the Gaussian scale mixture priors that are heleroscedastic across the matrix.

5. The apparatus of claim 2, wherein the Gaussian scale mixture priors have variance values that include (1) a row-dependent variance component that depends upon the matrix row and (2) a column-dependent variance component that depends upon the matrix column.

6. The apparatus of claim 1, wherein the BMF adds noise that varies with both row and column position in the matrix.

7. The apparatus of claim 1, further comprising:

an online retail store website embodied by one or more Internet-based server computers, the online retail store website displaying the recommendation to the user via an electronic device having a display that is operatively connected with the online retail store website via the Internet.

8. The apparatus of claim 7, wherein the electronic data processing device is said one or more Internet-based server computers.

9. A storage medium storing instructions executable by an electronic data processing device to perform a recommender method including:

(i) performing Bayesian Matrix Factorization (BMF) on a matrix having user and item dimensions using priors having noise terms that vary with both row and column position in the matrix to train a probabilistic collaborative filtering model and (ii) generating a recommendation for a user comprising a predicted item rating or identification of one or more recommended items using the probabilistic collaborative filtering model.

10. The storage medium of claim 9, wherein the priors comprise Gaussian priors.

11. The storage medium of claim 9, wherein the priors comprise non-Gaussian priors.

12. The storage medium of claim 12, wherein the non-Gaussian priors comprise Gaussian scale mixture priors.

14. The storage medium of claim 12, wherein the non-Gaussian priors are Student-t priors.

13. The storage medium of claim 13, wherein the hetersedastic noise terms comprise variance values of the Gaussian scale mixture priors that include (1) a row-dependent variance component that depends upon the matrix row and (2) a column-dependent variance component that depends upon the matrix column.

15. A method comprising:

performing Bayesian Matrix Factorization (BMF) on a matrix having user and item dimensions and matrix elements containing user ratings for items made by users wherein the BMF employs priors that include (1) a row-dependent variance component that depends upon the matrix row and (2) a column-dependent variance component that depends upon the matrix column to train probabilistic collaborative filtering model; and

generating a recommendation for a user comprising a predicted item rating or identification of one or more recommended items using the probabilistic collaborative filtering model;

wherein the performing and the generating are performed by an electronic data processing device.

16. The method of claim 16, wherein the priors comprise non-Gaussian priors.

18. The method of claim 16, wherein the priors comprise Student-t priors.

19. The method of claim 16, wherein one of:

(i) the rows of the matrix correspond to users and the row-dependent variance component depends on the number of user ratings made by the user corresponding to the row, and

(ii) the columns of the matrix correspond to users and the column-dependent variance component depends on the number of user ratings made by the user corresponding to the column.

20. The method of claim 16 wherein one of:

(i) the columns of the matrix correspond to items and the column-dependent variance component depends on the number of user ratings for the item corresponding to the column, and

(ii) the rows of the matrix correspond to items and the row-dependent variance component depends on the number of user ratings for the item corresponding to the row.

17. The method of claim 17, wherein the non-Gaussian priors comprise Gaussian scale mixture priors.