Generating item clusters based on aggregated search history data

• US 9,727,906 B1
• Filed: 12/15/2014
• Issued: 08/08/2017
• Est. Priority Date: 12/15/2014
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Sarmento, Luis Antonio Diniz Fernandes de Morais, Yu, Tao, Hinegardner, Lisa Jane, Husain, Haris, Spinelli, Alexander Michael
• Primary Examiner(s): Rosen, Nicholas D
• Application Number:
  US14/570,406
• Time in Patent Office: 967 Days
• Field of Search: None
• CPC Class Codes: G06F 16/355 Class or cluster creation or modification
• CPC Class Codes: G06F 16/957 Browsing optimisation, e.g. caching or content distillation
• CPC Class Codes: G06Q 30/0255 based on user history
• CPC Class Codes: G06Q 30/0256 User search
• CPC Class Codes: G06Q 30/0601 Electronic shopping [e-shopping]
• CPC Class Codes: G06Q 30/0631 Item recommendations

Abstract

The present disclosure provides computer-implemented systems and processes for clustering items and improving the utility of item recommendations. One process involves applying a clustering algorithm to users'"'"' search session queries over periods of time to generate query clusters comprising correlated query terms. Correlations may be based on, among other things, the frequency of which query term pairs appear together in a single search session. The generated query clusters may be used to generate item descriptor clusters indicative of items and/or types of items that may be complementary. Other criteria may be applied to the query and item clusters to generate variant clusters. For example, information such as related brands, market segments, and other data may be applied to item descriptor clusters to generate item clusters that include complementary items associated with or targeted for particular demographics. Item descriptor clusters and item clusters can be used to improve the item recommendations.

Claims

1. A computer-implemented method comprising:

under control of a computing device configured with specific computer-executable instructions;

accessing, from an electronic data store, search history data for a plurality of users of an electronic catalog, said search history data comprising session query logs, each session query log including one or more query terms associated with a search request submitted by a respective user;

analyzing the session query logs to generate one or more search session segments, each search session segment corresponding to a distinct search session and comprising one or more search requests;

for each respective search session segment, generating one or more query pairs, each query pair comprising at least two query terms that are associated with the one or more search requests comprised in the respective search session segment;

determining, for each respective query pair across all of the one or more search session segments, a query pair frequency indicative of how many times the respective query pair appears across all of the one or more search session segments;

clustering the query pairs, based at least in part on the determined query pair frequencies, into a query cluster, said query cluster comprising two or more query terms determined to be highly correlated based at least in part on the clustering;

accessing correlated attribute data, said correlated attribute data including two or more correlated attributes that tend to be correlated to each other;

generating an item descriptor cluster comprising combinations of the two or more query terms and the two or more correlated attributes;

generating, by executing a search request on the electronic catalog using the item descriptor cluster, an item cluster comprising items having at least one of the two or more correlated attributes; and

providing an item recommendation for at least one item contained in the item cluster.

2. The computer-implemented method of claim 1, wherein the item recommendation is provided in a search results page that is generated in response to receiving a current search request.

3. The computer-implemented method of claim 1, wherein the search history data comprises search history data gathered over a period of years.

4. The computer-implemented method of claim 1, wherein the session query logs collectively comprise at least millions of search requests received from at least millions of users.

5. The computer-implemented method of claim 1, wherein the clustering is based at least in part on at least one of Bayesian NMF clustering, K-means clustering, latent semantic analysis, or a cosine similarity technique.

6. A system comprising:

an electronic data store configured to at least store search history data associated with search requests submitted to an electronic catalog; and

a computing system comprising one or more hardware computing devices, said computing system in communication with the electronic data store and configured to at least;

analyze a plurality of submitted query terms associated with search session query logs to generate a plurality of query term pairs, each query term pair comprising at least two query terms that are associated with respective search requests associated with a single search session;

determine, for each respective query pair, a query pair frequency indicative of how often the query pair appears in the search session query logs;

cluster, based at least in part on the determined query pair frequencies, the query pairs into a query cluster, said query cluster comprising two or more query terms determined to be highly correlated based at least in part on said clustering;

access related item attribute data, said related item attribute data including two or more correlated item descriptors for related items, wherein the two or more correlated item descriptors tend to be correlated to each other;

generate an item descriptor cluster comprising combinations of the two or more query terms comprised in the query cluster and the two or more correlated item descriptors;

generate, by executing a search request on the electronic catalog using the item descriptor cluster, an item cluster comprising two or more items, each item of the two or more items having an attribute corresponding to at least one of the correlated item descriptors; and

providing an item recommendation comprising at least one of the two or more items.

7. The system of claim 6, wherein the item recommendation is provided in a catalog page that is generated by the electronic catalog.

8. The system of claim 6, wherein the session query logs collectively comprise at least millions of search requests.

9. The system of claim 6, wherein the clustering is based at least in part on at least one of Bayesian NMF clustering, K-means clustering, latent semantic analysis, or a cosine similarity technique.

10. The system of claim 6, wherein the related item attribute data includes at least one of:

related brands, prices, or target market segments.

11. A computer-implemented method comprising:

generating, based at least in part on a co-occurrence analysis of search histories associated with an electronic catalog, a cluster of item descriptors describing types of items in the electronic catalog that tend to be searched for in the electronic catalog in combination, the cluster of item descriptors including at least a first item descriptor representing a first item type and a second item descriptor representing a second item type;

generating, based at least in part on the cluster of item descriptors, a plurality of item clusters, each item cluster including a plurality of items in the electronic catalog and including at least a first item of the first item type and a second item of the second item type; and

providing a catalog item recommendation based at least in part on the plurality of item clusters,

said method performed by an electronic catalog system comprising one or more computing devices.

12. The computer-implemented method of claim 11, wherein the catalog item recommendation is provided in a search results page that is generated in response to receiving a current search request.

13. The computer-implemented method of claim 11, wherein generating the cluster of item descriptors further comprises:

generating a query cluster based at least in part on the co-occurrence analysis; and

generating the cluster of item descriptors based at least in part on the query cluster.

14. The computer-implemented method of claim 11, wherein the search histories collectively comprise at least millions of search requests.

15. The computer-implemented method of claim 11, wherein the co-occurrence analysis is performed using at least one of Bayesian NMF clustering, K-means clustering, latent semantic analysis, or a cosine similarity technique.

16. Non-transitory physical computer storage comprising computer-executable instructions stored thereon that, when executed by a hardware processor, are configured to perform operations comprising:

accessing, from an electronic data store that stores search session query log data, query pair frequency data indicative of how often a plurality of query pairs appear in search session logs associated with an electronic catalog;

generating a query cluster based at least in part on the accessed query pair frequency data by clustering the query pairs, said query cluster comprising at least two query terms determined to be highly correlated based at least in part on said clustering;

generating an item descriptor cluster comprising two or more query terms associated with the query cluster;

executing a search request on the electronic catalog using the item descriptor cluster to generate an item cluster comprising at least one item having an attribute corresponding to at least one of the two or more query terms; and

providing an item recommendation comprising the at least one item.

17. The non-transitory physical computer storage of claim 16, wherein the operations further comprise:

two or more correlated item attribute terms for related item attributes, wherein the two or more correlated item attributed terms tend to be correlated to each other; and

updating the item descriptor cluster to comprise combinations of the two or more query terms and the two or more correlated item attribute terms,

wherein the item cluster that is updated comprises at least one item having an attribute corresponding to at least one of the related item attributes.

18. The non-transitory physical computer storage of claim 16, wherein the item recommendation is provided in a search results page that is generated in response to receiving a current search request.

19. The non-transitory physical computer storage of claim 16, wherein the session query logs are gathered over a period of years.

20. The non-transitory physical computer storage of claim 16, wherein the session query logs collectively comprise at least millions of search requests received.

21. The non-transitory physical computer storage of claim 16, wherein the clustering is based at least in part on at least one of Bayesian NMF clustering, K-means clustering, latent semantic analysis, or a cosine similarity technique.