US 11,636,291 B1Granted

Content similarity determination

AssigneeAmazon Technologies, Inc.

InventorsZhen Zuo, Lixi Wang, Jianfeng Lin, Yi Sun, Michinari Momma, Yikai Ni, Wenbo Wang, Xiangdong Qian, Deqiang Meng

Life of this patent

Estimating the term…

Loading patent…

Abstract

Systems and techniques are generally described for determining similar content. First embedding data may be generated for first content based at least in part on a specified similarity task. A first query comprising the first embedding data may be generated and sent to a first database partition. A first search result representing second content may be determined using the first embedding data. A second query comprising the first embedding data may be generated and sent to a second database partition different from the first data base partition. A second search result representing third content may be determined using the first embedding data. In some examples, output data comprising at least one of the first search result and the second search result may be generated. The output data may represent content classified as similar to the first content.

Claims20

Dependent claims

A method of determining similarity between content, comprising:

generating, for first content, first embedding data representing the first content, wherein the first embedding data is generated using an image feature vector, a title feature vector, a description feature vector, and a user behavior feature vector related to user interaction with the first content;
generating a plurality of queries comprising the first embedding data;
sending a first query of the plurality of queries to a first database partition;
receiving a first query result from the first database partition, wherein the first query result comprises second content associated with second embedding data that is a first closest distance from the first embedding data in an embedding space shared by the first embedding data and the second embedding data;
sending a second query of the plurality of queries to a second database partition;
receiving a second query result from the second database partition, wherein the second query result comprises third content associated with third embedding data that is a second closest distance from the first embedding data in the embedding space;
determining filter attribute data representing a first content attribute;
determining that the second content includes the first content attribute and that the third content does not include the first content attribute;
and
generating result data comprising the second content, wherein the result data represents content classified as similar to the first content.

7 more elements

A system comprising:

at least one processor;
and
at least one non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to cause the at least one processor to:
receive first data identifying a first similarity task from among a plurality of similarity tasks;
determine a first set of feature data corresponding to the first similarity task, the first set of feature data comprising at least one image feature and at least one text feature;
generate first embedding data for first content based at least in part on the first data, wherein the first embedding data is generated using the first set of features;
generate a first query comprising the first embedding data;
send the first query to a first database partition;
determine, based at least in part on the first embedding data, a first search result representing second content;
generate a second query comprising the first embedding data;
send the second query to a second database partition different from the first database partition;
determine, based at least in part on the first embedding data, a second search result representing third content;
and
generate output data comprising at least one of the first search result and the second search result, the output data representing content classified as similar to the first content.

11 more elements

Dependencies are inferred from the claim text.

Facts

Application: 16/996,528Filed 18 Aug 2020 · us provisional application
Priority: 6 Apr 2020
Granted: 25 Apr 2023
Term ends: Not estimatedThe source records don’t support an estimate.

Classification17 codes

G06F 18/24147
Based on distances to training or reference patterns: distances to closest patterns, e.g. nearest neighbour classification
G06N 3/084
Learning methods: backpropagation, e.g. using gradient descent
G06F 16/41
Of multimedia data, e.g. slideshows comprising image and additional audio data: indexing; Data structures therefor; Storage structures
G06F 16/434
Query formulation: using image data, e.g. images, photos, pictures taken by a user