Back to results
US 11,106,859 B1Granted
Systems and methods for page embedding generation
AssigneeFacebook, Inc.
InventorsYixin Li, Wenzhe Jiang, Jing He, Jinyi Yao
Life of this patent
Estimating the term…
Abstract
Systems, methods, and non-transitory computer readable media are configured to generate multiple channel embeddings for a page of a social networking system. The multiple channel embeddings can be mapped to a shared embedding space. A page embedding for the page of the social networking system can then be generated.
Claims20
A computer-implemented method comprising:
- generating, by a computing system, a plurality of embedding sequences for a page based at least in part on sequences of tokens corresponding to a plurality of channel types, wherein the plurality of channel types includes a page hashtag channel type, and the tokens corresponding to the page hashtag channel type include hashtags generated from images on the page;
- generating, by the computing system, a plurality of channel embeddings associated with a plurality of channels on the page corresponding to the plurality of channel types, wherein a channel embedding of the plurality of channel embeddings is based at least in part on a first pooling of a corresponding embedding sequence of the plurality of embedding sequences;
- mapping, by the computing system, the plurality of channel embeddings to a shared embedding space;
- generating, by the computing system, a page embedding for the page based at least in part on a second pooling of the plurality of channel embeddings;
- determining, by the computing system, a loss value that is related to a first similarity value between the page embedding and a random page embedding of a random page and that is inversely related to a second similarity value between the page embedding and a related page embedding of a related page, wherein weights associated with at least one of the generating the plurality of embedding sequences, the generating the plurality of channel embeddings, or the generating the page embeddings are updated based on the loss value;
- and
- determining, by the computing system, a likelihood a user will interact with the page based at least in part on the page embedding and a user embedding for the user.
4 more elements
A system comprising:
- at least one processor;
- and
- a memory storing instructions that, when executed by the at least one processor, cause the system to perform:
- generating a plurality of embedding sequences for a page based at least in part on sequences of tokens corresponding to a plurality of channel types, wherein the plurality of channel types includes a page hashtag channel type, and the tokens corresponding to the page hashtag channel type include hashtags generated from images on the page;
- generating a plurality of channel embeddings associated with a plurality of channels on the page corresponding to the plurality of channel types, wherein a channel embedding of the plurality of channel embeddings is based at least in part on a first pooling of a corresponding embedding sequence of the plurality of embedding sequences;
- mapping the plurality of channel embeddings to a shared embedding space;
- generating a page embedding for the page based at least in part on a second pooling of the plurality of channel embeddings;
- determining a loss value that is related to a first similarity value between the page embedding and a random page embedding of a random page and that is inversely related to a second similarity value between the page embedding and a related page embedding of a related page, wherein weights associated with at least one of the generating the plurality of embedding sequences, the generating the plurality of channel embeddings, or the generating the page embeddings are updated based on the loss value;
- and
- determining a likelihood a user will interact with the page based at least in part on the page embedding and a user embedding for the user.
7 more elements
A non-transitory computer-readable storage medium including instructions that, when executed by at least one processor of a computing system, cause the computing system to perform:
- generating a plurality of embedding sequences for a page based at least in part on sequences of tokens corresponding to a plurality of channel types, wherein the plurality of channel types includes a page hashtag channel type, and the tokens corresponding to the page hashtag channel type include hashtags generated from images on the page;
- generating a plurality of channel embeddings associated with a plurality of channels on the page corresponding to the plurality of channel types, wherein a channel embedding of the plurality of channel embeddings is based at least in part on a first pooling of a corresponding embedding sequence of the plurality of embedding sequences;
- mapping the plurality of channel embeddings to a shared embedding space;
- generating a page embedding for the page based at least in part on a second pooling of the plurality of channel embeddings;
- determining a loss value that is related to a first similarity value between the page embedding and a random page embedding of a random page and that is inversely related to a second similarity value between the page embedding and a related page embedding of a related page, wherein weights associated with at least one of the generating the plurality of embedding sequences, the generating the plurality of channel embeddings, or the generating the page embeddings are updated based on the loss value;
- and
- determining a likelihood a user will interact with the page based at least in part on the page embedding and a user embedding for the user.
4 more elements
Dependencies are inferred from the claim text.