Recommender system with training function based on non-random missing data

• US 9,092,739 B2
• Filed: 07/22/2010
• Issued: 07/28/2015
• Est. Priority Date: 07/22/2010
• Status: Active Grant

• Current Assignee: Alcatel-Lucent SA (Nokia Corporation)
• Original Assignee: Alcatel-Lucent SA (Nokia Corporation)
• Inventors: Steck, Harald
• Primary Examiner(s): Chaki, Kakali
• Assistant Examiner(s): HANCHAK, WALTER P
• Application Number:
  US12/841,720
• Publication Number:
  US 20120023045A1
• Time in Patent Office: 1,832 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

A processing device of an information processing system is operative to obtain observed feedback data, to construct a model that accounts for both the observed feedback data and additional feedback data that is missing from the observed feedback data, to optimize one or more parameters of the model using a training objective function, and to generate a list of recommended items for a given user based on the optimized model. In illustrative embodiments, the missing feedback data comprises data that is missing not at random (MNAR), and the model comprises a matrix factorization model. The processing device may implement a recommender system comprising a training module coupled to a recommendation module.

Claims

1. A method for use in a recommender system, the method comprising:

obtaining observed feedback data;

constructing a model that accounts for both the observed feedback data and additional feedback data that is missing from the observed feedback data;

optimizing one or more parameters of the model using a training objective function; and

generating a list of recommended items for a given user based on the optimized model;

wherein the constructing step comprises;

determining whether each of a plurality of item-user pairs are associated with at least one of the observed feedback data and the additional feedback data that is missing from the observed feedback data; and

responsive to said determining, assigning the plurality of item-user pairs to respective ones of a plurality of different classes;

wherein the training objective function utilizes weights associated with respective ones of the item-user pairs, the weights assigned to the item-user pairs being based at least in part on the classes of the item-user pairs;

wherein the training objective function comprises at least one of;

a penalized log likelihood logistic regression using the weights for item-user pairs, the weights being binary weights;

a least squares regression using the weights for item-user pairs, the weights being binary weights;

a regression using the weights for item-user pairs, the regression being based at least in part on a comparison of observed or imputed ratings to predicted ratings for one or more item-user pairs; and

a regression using the weights for item-user pairs, the regression being based at least in part on a given item-user pair having two or more observed or imputed ratings; and

wherein the obtaining, constructing, optimizing and generating steps are implemented in a processing device comprising a processor coupled to a memory.

2. The method of claim 1 wherein the missing feedback data comprises data that is missing not at random (MNAR).

3. The method of claim 1 wherein the observed feedback data comprises explicit feedback data including ratings assigned to selectable items by users.

4. The method of claim 1 wherein the model comprises a matrix factorization model.

5. The method of claim 4 wherein the matrix factorization model is characterized as a neural network.

6. The method of claim 1 wherein the training objective function is given by:

7. The method of claim 1 wherein the training objective function is given by:

8. The method of claim 1 wherein the training objective function is given by:

9. The method of claim 8 wherein the weights W_i,u are determined in the following manner:

10. The method of claim 1 wherein the ,training objective function is given by:

11. An article of manufacture comprising a non-transitory computer-readable storage medium having embodied therein executable program code that when executed by a processor of a processing device causes the device to perform the steps of the method of claim 1.

21. The method of claim 1 wherein the training objective function comprises the penalized log likelihood logistic regression using the weights for item-user pairs, the weights being binary weights.

22. The method of claim 1 wherein the training objective function comprises the least squares regression using the weights for item-user pairs, the weights being binary weights.

23. The method of claim 1 wherein the training objective function comprises the regression using the weights for item-user pairs, the regression being based at least in part on the comparison of observed or imputed ratings to predicted ratings for one or more item-user pairs.

24. The method of claim 1 wherein the training objective function comprises the regression using the weights for item-user pairs, the regression being based at least in part on the given item-user pair having two or more observed or imputed ratings.

12. An apparatus comprising:

a processing device comprising a processor having an associated memory;

wherein the processing device is operative;

to obtain observed feedback data;

to construct a model that accounts for both the observed feedback data and additional feedback data that is missing from the observed feedback data;

to optimize one or more parameters of the model using a training objective function; and

to generate a list of recommended items for a given user based on the optimized model;

wherein constructing the model comprises;

determining whether each of a plurality of item-user pairs are associated with at least one of the observed feedback data and the additional feedback data that is missing from the observed feedback data; and

responsive to said determining, assigning the plurality of item-user pairs to respective ones of a plurality of different classes; and

wherein the training objective function utilizes weights associated with respective ones of the item-user pairs, the weights assigned to the item-user pairs being based at least in part on the classes of the item-user pairs; and

wherein the training objective function comprises at least one of;

a penalized log likelihood logistic regression using the weights for item-user pairs, the weights being binary weights;

a least squares regression using the weights for item-user pairs, the weights being binary weights;

a regression using the weights for item-user pairs, the regression being based at least in part on a comparison of observed or imputed ratings to predicted ratings for one or more item-user pairs; and

a regression using the weights for item-user pairs, the regression being based at least in part on a given item-user pair having two or more observed or imputed ratings.

13. The apparatus of claim 12 wherein the processing device comprises a server.

14. The apparatus of claim 12 wherein the server comprises a back end server of a web-based information processing system.

15. The apparatus of claim 12 wherein the processing device implements a recommender system comprising a recommendation module and a training module.

16. The apparatus of claim 12 wherein the apparatus is comprised in an integrated circuit.

17. A recommender system comprising:

a training module; and

a recommendation module coupled to the training module;

wherein the training module is configured to apply a training objective function to optimize one or more parameters of a model that accounts for both observed feedback data and additional feedback data that is missing from the observed feedback data;

wherein the recommendation module generates a list of recommended items for a given user based on the optimized model;

wherein the model is constructed at least in part by;

determining whether each of a plurality of item-user pairs are associated with at least one of the observed feedback data and the additional feedback data that is missing from the observed feedback data; and

responsive to said determining, assigning the plurality of item-user pairs to respective ones of a plurality of different classes;

wherein the training objective function utilizes weights associated with respective ones of the item-user pairs, the weights assigned to the item-user pairs being based at least in part on the classes of the item-user pairs;

wherein the training objective function comprises at least one of;

a penalized log likelihood logistic regression using the weights for item-user pairs, the weights being binary weights;

a least squares regression using the weights for item-user pairs, the weights being binary weights;

a regression using the weights for item-user pairs, the regression being based at least in part on a comparison of observed or imputed ratings to predicted ratings for one or more item-user pairs; and

a regression using the weights for item-user pairs, the regression being based at least in part on a given item-user pair having two or more observed or imputed ratings; and

wherein the recommender system is implemented using at least one processor device.

18. The system of claim 17 wherein the missing feedback data comprises data that is missing not at random (MNAR).

19. The system of claim 17 wherein the observed feedback data comprises explicit feedback data including ratings assigned to selectable items by users.

20. The system of claim 17 wherein the model comprises a matrix factorization model.