Processes for improving the utility of personalized recommendations generated by a recommendation engine

• US 7,921,071 B2
• Filed: 11/16/2007
• Issued: 04/05/2011
• Est. Priority Date: 11/16/2007
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Hicks, Cory
• Primary Examiner(s): Sparks; Donald
• Assistant Examiner(s): Bharadwaj; Kalpana
• Application Number:
  US11/941,809
• Publication Number:
  US 20090132459A1
• Time in Patent Office: 1,236 Days
• Field of Search: 702/189, 725/46
• US Class Current: 706/52
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation
• CPC Class Codes: G06Q 30/0269 based on user profile or attribute

Abstract

A recommendation system increases the diversity of item recommendations provided to a target user by using item similarity data to filter an initial recommendation set. In one embodiment, selected items are filtered from the initial recommendation set based on similarity scores that represent degrees of similarity between particular items. The similarity scores may be generated based on an automated comparison of item attributes or content, or based on another measure of item similarity.

Claims

1. A computer-implemented method, comprising:

retrieving item preference data for a target user from computer storage, said item preference data reflective of item preferences of the target user;

generating a recommendation set using the item preference data, said recommendation set comprising a computer representation of items predicted to be of interest to the target user;

generating a filtered recommendation set by using item similarity data to filter the recommendation set, said filtered recommendation set having a greater degree of item diversity than the recommendation set, wherein generating the filtered recommendation set comprises identifying a fuzzy duplicate item pair in the recommendation set, and removing one item of said fuzzy duplicate item pair from the recommendation set; and

outputting information regarding at least some of the items in the filtered recommendation set for presentation to the target user.

2. The method of claim 1, wherein generating the filtered recommendation set comprises identifying an item pair for which an item similarity score satisfies a threshold, and removing one item of said item pair from the recommendation set.

3. The method of claim 1, wherein generating the filtered recommendation set comprises removing a first item from the recommendation set in response to detecting that the first item has a threshold degree of similarity to a second item in the recommendation set.

4. The method of claim 1, wherein generating the filtered recommendation set comprises preventing two items that are fuzzy duplicates of each other from being recommended to the target user on a common display page.

5. The method of claim 1, wherein generating the filtered recommendation set comprises applying filtering only to a selected subset of the items in the recommendation set.

6. The method of claim 1, wherein generating the filtered recommendation set comprises using attribute-based item similarity scores to select items to filter from the recommendation set.

7. The method of claim 1, wherein the recommendation set is generated based on behavioral associations between the items, and the filtered recommendation set is generated using data regarding attribute-based similarities between the items.

8. The method of claim 1, wherein the recommendation set is generated by a recommendation engine, and the step of generating the filtered recommendation set is performed by a filtering component that is separate from the recommendation engine.

9. The method of claim 1, wherein the step of outputting information regarding at least some of the items comprises generating and transmitting a web page that includes a display element that enables a user to cause one or more filtered out items to be displayed.

10. A computer system comprising:

one or more computers programmed with executable code modules to at least;

retrieve item preference data for a target user from computer storage, said item preference data reflective of item preferences of the target user;

generate a recommendation set using the item preference data, said recommendation set comprising a computer representation of items predicted to be of interest to the target user;

generate a filtered recommendation set by using item similarity data to filter the recommendation set, said filtered recommendation set having a greater degree of item diversity than the recommendation set, wherein generating the filtered recommendation set comprises identifying a fuzzy duplicate item pair in the recommendation set, and removing one item of said fuzzy duplicate item pair from the recommendation set; and

output information regarding at least some of the items in the filtered recommendation set for presentation to the target user.

11. A computer-implemented method, comprising:

receiving a recommendation set generated by a recommendation engine, said recommendation set comprising a computer representation of a plurality of items predicted to be of interest to a target user; and

filtering the recommendation set to prevent at least some items that have a threshold degree of similarity to each other from being recommended to the target user in combination, said threshold set so as to prevent, for at least some fuzzy duplicate items pairs, both items of the pair from being recommended in combination, said filtering performed by execution of program code by one or more processors.

12. The method of claim 11, wherein filtering the recommendation set comprises using pre-generated item similarity data to identify item pairs that satisfy said threshold.

13. The method of claim 11, further comprising generating a web page for presentation to a user, said web page including a description of an item in the filtered recommendation set, and including a display element that is selectable to view at least one corresponding filtered-out item.

14. The method of claim 11, wherein filtering the recommendation set comprises, after receiving said recommendation set, calculating distances between items in the recommendation set, and using the distances to identify item pairs that satisfy said threshold.

15. The method of claim 11, wherein filtering the recommendation set comprises using a probabilistic selection algorithm to select items to filter out.

16. The method of claim 11, wherein the recommendation set is received via a call to a web service.

17. A computer-readable medium which stores a computer program comprising executable instructions that direct a computer system to at least:

receive a recommendation set generated by a recommendation engine, said recommendation set comprising a computer representation of a plurality of items predicted to be of interest to a target user; and

filter the recommendation set to prevent at least some items that have a threshold degree of similarity to each other from being recommended to the target user in combination, said threshold set so as to prevent, for at least some fuzzy duplicate items pairs, both items of the pair from being recommended in combination.

18. A recommendation system, comprising:

a recommendation engine configured to generate personalized item recommendations for target users based on item preference data stored for said target users; and

a similar items filter configured to filter the personalized item recommendations generated by the recommendation engine based on item similarities data to prevent at least some items that are fuzzy duplicates of each other from being recommended in combination to a target user, said similar items filter thereby increasing a diversity of personalized recommendations presented to target users;

said recommendation engine and similar items filter comprising computer hardware that executes software.

19. The recommendation system of claim 18, wherein the similar items filter is configured to use pre-generated item similarities data to filter the personalized item recommendations.

20. The recommendation system of claim 18, further comprising a page generator configured to generate recommendation pages, said page generator responsive to removal by the similar items filter of a first item of an item pair by incorporating a display element into a recommendation page, said display element provided in association with a description of a second item of said item pair, and enabling a user to view a description of at least the first item.

21. The recommendation system of claim 18, wherein the recommendation system is configured to use behavior-based associations between items to generate the personalized item recommendations, and the similar items filter is configured to use attribute-based similarities between the items to filter the personalized recommendations.

22. The recommendation system of claim 18, wherein the similar items filter uses a similarity threshold to identify items to filter from recommendation sets, said similarity threshold set so as to identify fuzzy duplicate item pairs.