Identifying similar content in a multi-item embedding space

Patent metadata

Field	Value
Publication number	US-12475171-B2
Application number	US-202318163716-A
Country	US
Kind code	B2
Filing date	Feb 2, 2023
Priority date	Dec 12, 2019
Publication date	Nov 18, 2025
Grant date	Nov 18, 2025

Abstract

Systems and methods for identifying content for an input query are presented. A mapping model is trained to map elements of an input query embedding vector for a received query into one or more elements of a destination embedding vector. In response to receiving an input query, an input query embedding vector is generated that projects into an input query embedding space. The input query embedding vector is processed by the mapping model to map the input query embedding vector into one or more elements of a destination embedding vector in a destination embedding space, resulting in a partial destination embedding vector. Items of a corpus of content are projected into the destination embedding space and the partial destination embedding vector is also projected into the destination embedding space. A similarity measure determines the most-similar items to the partial destination embedding vector and at least some of the most-similar items are returned in response to the input query.

First claim

What is claimed:

1 . A computer-implemented method, comprising:
receiving a query including a first plurality of content items, wherein the first plurality of content items includes content items of different content types; generating a query embedding vector for the first plurality of content items representing an entirety of the first plurality of content items, the generating the query embedding vector comprising:
generating a single-item embedding vector from the query, the single-item embedding vector having a first number of elements; and
mapping the first number of elements to elements of a partial multi-item embedding vector corresponding to the query embedding vector, the partial multi-item embedding vector being configured for projection to a multi-item embedding space;
projecting the query embedding vector into a multi-item embedding space; determining a plurality of similarity measures between the query embedding vector and a first plurality of multi-item embedding vectors, wherein:
the first plurality of multi-item embedding vectors are projected into the multi-item embedding space, each multi-item embedding vector having elements corresponding to distinct items of a corresponding content group, each of a plurality of content groups including a plurality of items; and
identifying a second plurality of multi-item embedding vectors from the first plurality of multi-item embedding vectors based at least in part on the plurality of similarity measures; and providing at least some of the respective content groups corresponding to the second plurality of multi-item embedding vectors as a response to the query.

2 . The computer-implemented method of claim 1 , wherein the query includes a user post including a collection of the first plurality of content items.

3 . The computer-implemented method of claim 1 , wherein each content group includes content items of different content types, and wherein the different content types include any two or more of audio content types, video content types, image content types, textual content, uniform resource locator (URL) source information, and a filename.

4 . The computer-implemented method of claim 1 , wherein:
the first plurality of multi-item embedding vectors are sectioned into a plurality of sections according to a Locality Sensitive Hashing (LSH) technique; and identifying the second plurality of multi-item embedding vectors from the first plurality of multi-item embedding vectors includes determining a first section of the plurality of sections.

5 . The computer-implemented method of claim 4 , wherein determining the first section of the plurality of sections is based at least in part on a hashing of the query embedding vector into the first section of the plurality of sections.

6 . A computing system, comprising:
one or more processors; and a memory storing program instructions that, when executed by the one or more processors, cause the one or more processors to at least:
receive a query embedding vector representing a query, the generating the query embedding vector comprising generating a single-item embedding vector from the query, the single-item embedding vector having a first number of elements;
determine a content type associated with the query;
determine, based on the content type, a trained mapping model trained to map an input embedding vector to an output multi-item embedding vector;
map, using the trained mapping model, the query embedding vector to a multi-item query embedding vector, the mapping comprising mapping the first number of elements to elements of a partial multi-item embedding vector corresponding to the query embedding vector, the partial multi-item embedding vector being configured for projection to a multi-item embedding space;
project the multi-item query embedding vector into the multi-item embedding space;
identify a first plurality of multi-item embedding vectors from a second plurality of multi-item embedding vectors based at least in part on a plurality of similarity measures between the multi-item query embedding vector and the second plurality of multi-item embedding vectors, wherein:
the second plurality of multi-item embedding vectors are projected into the multi-item embedding space, each multi-item embedding vector of the second plurality of multi-item embedding vectors having elements corresponding to distinct items of a corresponding content group, each of a plurality of content groups including a plurality of items; and

providing at least a portion of the plurality of content groups corresponding to the first plurality of multi-item embedding vectors as a response to the query.

7 . The computing system of claim 6 , wherein mapping of the query embedding vector includes mapping values of the query embedding vector to the multi-item query embedding vector to form a partial multi-item query embedding vector.

8 . The computing system of claim 7 , wherein a plurality of elements of the query embedding vector is mapped to a single element of the multi-item query embedding vector.

9 . The computing system of claim 7 , wherein at least one element of the query embedding vector does not include a corresponding value in the multi-item query embedding vector.

10 . The computing system of claim 6 , wherein the content type associated with the query includes a first plurality of content types.

11 . The computing system of claim 10 , wherein:
the multi-item query embedding vector is associated with a second plurality of content types; and mapping of the query embedding vector to the multi-item query embedding vector includes:
determining overlapping content types between the first plurality of content types and the second plurality of content types; and
mapping features of the query embedding vector corresponding to the overlapping content types to the multi-item query embedding vector.

12 . The computing system of claim 6 , wherein:
the second plurality of multi-item embedding vectors are sectioned into a plurality of sections; and identification of the first plurality of multi-item embedding vectors from the second plurality of multi-item embedding vectors includes:
associating the multi-item query embedding vector with a first section of the plurality of sections; and
determining a third plurality of multi-item embedding vectors associated with the first section as the first plurality of multi-item embedding vectors.

13 . The computing system of claim 12 , wherein the second plurality of multi-item embedding vectors are sectioned in accordance with a Locality Sensitive Hashing (LSH) technique.

14 . The computing system of claim 6 , wherein the plurality of similarity measures includes a cosine similarity measure.

15 . The computing system of claim 6 , wherein each of the plurality of content groups corresponds to a user post including a collection of content items of different content types.

16 . A computer-implemented method, comprising:
receiving a query embedding vector associated with a query, wherein the query is associated with a single content type, and wherein the query embedding vector has a first plurality of values; generating, based at least in part on the query embedding vector, a partial multi-item query embedding vector, comprising:
mapping the first plurality of values to a second plurality of values of the partial multi-item query embedding vector, wherein the partial multi-item query embedding vector is a multi-item embedding vector with the second plurality of values corresponding to the single content type of the query

Assignees

• Pinterest Inc

Inventors

• Zhai Andrew Huan
• Chen Kaifeng
• Rosenberg Charles Joseph

Classifications

• G06F16/84

  Mapping; Conversion · CPC title

• G06F16/9032

  Query formulation · CPC title

• G06F16/958

  Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking · CPC title

• G06F16/906

  Clustering; Classification · CPC title

• G06F16/3347

  using vector based model · CPC title