US 12,019,671 B2
Locally constrained self-attentive sequential recommendation
Handong Zhao, Cupertino, CA (US); Zhankui He, San Diego, CA (US); Zhaowen Wang, San Jose, CA (US); Ajinkya Gorakhnath Kale, San Jose, CA (US); and Zhe Lin, Clyde Hill, WA (US)
Assigned to Adobe Inc., San Jose, CA (US)
Filed by Adobe Inc., San Jose, CA (US)
Filed on Oct. 14, 2021, as Appl. No. 17/501,191.
Prior Publication US 2023/0116969 A1, Apr. 20, 2023
Int. Cl. G06F 16/43 (2019.01); G06F 16/438 (2019.01); G06F 16/44 (2019.01); G06N 3/045 (2023.01)
CPC G06F 16/438 (2019.01) [G06F 16/447 (2019.01); G06N 3/045 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A system comprising:
a search query module implemented in hardware of a computing device, the search query module receiving a search query from an entity to search a plurality of digital content;
a sequence data module implemented in hardware of the computing device, the sequence data module obtaining sequence input data describing a sequence of the digital content, with which, the entity has interacted, the sequence of the digital content including a first item from the plurality of digital content with which the entity interacted and a second item from the plurality of digital content with which the entity interacted, the first item being from a first search performed by the entity and the second item being from a second search performed by the entity;
a sequence search module implemented in hardware of the computing device, the sequence search module generating a search result based on the search query and the sequence input data, the sequence search module including a machine-learning module having a localized self-attention module specifying a localized self-attention bias based on a temporal order of digital content within the sequence, the localized self-attention module including a masking-based local encoder implementing the localized self-attention bias, the self-attention bias assigning greater weight to the first item relative to the second item based upon the first item occurring more recently in the sequence than the second item;
a search result representation module implemented in hardware of the computing device, the search result representation module generating a representation of the first item included in the search result, the search result representation module using the greater weight assigned to the first item as a basis for including the representation of the first item in the search result; and
a search result output module implemented in hardware of the computing device, the search result output module outputting the search result having the representation.