Probabilistic Recommendation System

• US 20080294617A1
• Filed: 05/22/2007
• Published: 11/27/2008
• Est. Priority Date: 05/22/2007
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Smith, Brent, Chakrabarti, Kushal
• Granted Patent: US 8,301,623 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/9535 Search customisation based on user profiles and personalisation
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising

Abstract

A recommendations system uses probabilistic methods to select, from a candidate set of items, a set of items to recommend to a target user. Some embodiments of the methods effectively introduce noise into the recommendations process, causing the recommendations presented to the target user to vary in a controlled manner from one visit to the next. The methods may increase the likelihood that at least some of the items recommended over a sequence of visits will be useful to the target user. Some embodiments of the methods are stateless such that the system need not keep track of which items have been recommended to which users.

Claims

1. A computer-implemented method of selecting items to recommend, the method comprising:

generating a ranked set of candidate items to recommend to a target user;

probabilistically varying rankings of at least some of the candidate items in the ranked set to generate a probabilistically-varied ranked set of candidate items; and

selecting, from said probabilistically-varied ranked set, a subset of items to present to the user.

2. The computer-implemented method of claim 1, wherein selecting, from said probabilistically-varied ranked set, a subset of items to present to the user comprises selecting a most-highly ranked subset of items.

3. The computer-implemented method of claim 1, wherein at least some of the candidate items are associated with one or more identified reasons for being recommended to the target user.

4. The computer-implemented method of claim 3, wherein the degree to which the ranking of a candidate item is probabilistically varied depends on the number of identified reasons for recommending the candidate item to the target user.

5. The computer-implemented method of claim 4, wherein probabilistically varying the rankings comprises causing candidate items having relatively low numbers of reasons to be probabilistically varied to a greater degree than candidate items having relatively high numbers of reasons.

6. The computer-implemented method of claim 1, wherein probabilistically varying the rankings comprises causing candidate items having relatively low category scores to be probabilistically varied to a greater degree than candidate items having relatively high category scores.

7. The computer-implemented method of claim 1, wherein probabilistically varying the rankings comprises generating values for at least a portion of the candidate items and adding probabilistic noise to the values.

8. The computer-implemented method of claim 7, wherein adding probabilistic noise to the values comprises generating a probability distribution corresponding to each value, wherein the characteristics of the probability distribution depend on the corresponding value.

9. The computer-implemented method of claim 8, further comprising sampling values from each probability distribution to generate a score for said at least a portion of the candidate items, wherein the score is responsive to the sampled value.

10. The computer-implemented method of claim 1, wherein probabilistically varying rankings of at least some of the candidate items in the ranked set further comprises probabilistically varying the rankings to achieve an overlap of 5 to 20 percent of recommended items between recommendation generation events.

11. A computer-implemented method of selecting items to recommend, the method comprising:

identifying, for a candidate item, a plurality of reasons for recommending the candidate item to a target user, each reason corresponding to a different respective event or set of events recorded in an event history of the user;

assigning a score to the candidate item responsive to a value probabilistically selected according to a probability distribution, the probability distribution having at least one characteristic that depends on the number of said reasons; and

using the score to determine whether to recommend the item to the target user.

12. The computer-implemented method of claim 11, wherein the at least one characteristic of the probability distribution comprises entropy.

13. The computer-implemented method of claim 12, wherein the entropy of the probability distribution decreases as the number of reasons increases.

14. The computer-implemented method of claim 11, wherein the probability distribution is a normal distribution.

15. The computer-implemented method of claim 14, wherein the at least one characteristic of the probability distribution comprises a mean configured to depend on the number of reasons, and wherein the score is responsive to the mean.

16. The computer-implemented method of claim 14, wherein the at least one characteristic of the probability distribution comprises a standard deviation configured to depend on the number of reasons, and wherein the score is responsive to the standard deviation.

17. The computer-implemented method of claim 11, wherein the score depends upon strength values associated with specific reasons.

18. The computer-implemented method of claim 11, further comprising assigning a negative feedback score to the candidate item, wherein the negative feedback score is dependent on the number of said reasons.

19. The computer-implemented method of claim 18, wherein assigning a score to the candidate item further comprises determining one or more category scores responsive to the number of reasons in one or more categories selected from the group of a purchase history category, an unbought interests category, and a recent activity category.

20. The computer-implemented method of claim 19, wherein assigning a score to the candidate item further comprises adding the one or more category scores and the negative feedback score.

21. The computer-implemented method of claim 11, wherein assigning a score to the candidate item responsive to a value probabilistically selected according to a probability distribution further comprises assigning the score such that an overlap of 5 to 20 percent of recommended items between recommendation generation events is achieved.

22. A computer-implemented method, the method comprising:

assigning initial scores to candidate items generated as preliminary recommendations to a target user, the initial scores based on a number of reasons corresponding to different respective events recorded in an event history of the user;

adding probabilistic noise to at least some of the initial scores;

sampling values from a probability distribution that includes the probabilistic noise; and

generating adjusted scores for the at least some of the candidate items, the adjusted scores being responsive to the sampled values.

23. The computer-implemented method of claim 22, wherein sampling values from a probability distribution that includes the probabilistic noise comprises probabilistically selecting values according to a probability distribution having at least one characteristic that depends on the number of said reasons.

24. The computer-implemented method of claim 22, wherein the probability distribution is a normal distribution.

25. The computer-implemented method of claim 24, wherein a mean of the normal distribution depends on the number of reasons, and wherein the score is responsive to the mean.

26. The computer-implemented method of claim 24, wherein a standard deviation of the normal distribution depends on the number of reasons, and wherein the score is responsive to the standard deviation.

27. The computer-implemented method of claim 22, wherein generating adjusted scores for at least some of the candidate items further comprises determining a negative feedback score dependent on the number of reasons.

28. The computer-implemented method of claim 27, wherein assigning initial scores to candidate items further comprises determining one or more category scores responsive to the number of reasons in one or more categories selected from the group of a purchase history category, an unbought interests category, and a recent activity category.

29. The computer-implemented method of claim 28, wherein assigning initial scores to candidate items further comprises adding the one or more category scores and the negative feedback score.

30. The computer-implemented method of claim 22, wherein generating adjusted scores for the at least some of the candidate items further comprises generating adjusted scores such that an overlap of 5 to 20 percent of recommended items between recommendation generation events is achieved.

31. A recommendation engine for selecting items to recommend, the recommendation engine comprising:

a candidate generator configured to generate a ranked set of candidate items to recommend to a target user;

a probabilistic scorer configured to probabilistically vary rankings of at least some of the candidate items in the ranked set to generate a probabilistically-varied ranked set of candidate items; and

a candidate filter configured to select, from said probabilistically-varied ranked set, a subset of items to present to the user.

32. The recommendation engine of claim 31, wherein the degree to which the probabilistic scorer probabilistically varies the ranking of a candidate item depends on a number of identified reasons for recommending the candidate item to the target user.

33. The recommendation engine of claim 32, wherein probabilistically varying the rankings comprises causing candidate items having relatively low numbers of reasons to be probabilistically varied to a greater degree than candidate items having relatively high numbers of reasons.