Pinterest, Inc.

Amazon Technologies, Inc.

Kaifeng Chen

Jurij Leskovec

Chantat Eksombatchai

Ruining He

Rex Ying

Charles Joseph Rosenberg

Andrew Huan Zhai

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.

Identifying similar content in a multi-item embedding space

Systems and methods for generating embeddings for nodes of a corpus graph are presented. More particularly, operations for generation of an aggregated embedding vector for a target node is efficiently divided among operations on a central processing unit and operations on a graphic processing unit. With regard to a target node within a corpus graph, processing by one or more central processing units (CPUs) is conducted to identify the target node's relevant neighborhood (of nodes) within the corpus graph. This information is prepared and passed to one or more graphic processing units (GPUs) that determines the aggregated embedding vector for the target node according to data of the relevant neighborhood of the target node.

Generating neighborhood convolutions within a large network

Systems and methods for generating embeddings for nodes of a corpus graph are presented. The embeddings correspond to aggregated embedding vectors for nodes of the corpus graph. Without processing the entire corpus graph to generate all aggregated embedding vectors, a relevant neighborhood of nodes within the corpus graph are identified for a target node of the corpus graph. Based on embedding information of the target node's immediate neighbors, and also upon neighborhood embedding information from the target node's relevant neighborhood, an aggregated embedding vector can be generated for the target node that comprises both an embedding vector portion corresponding to the target node, as well as a neighborhood embedding vector portion, corresponding to embedding information of the relevant neighborhood of the target node. Utilizing both portions of the aggregated embedding vector leads to improved content recommendation to a user in response to a query.

Efficient convolutional network for recommender systems

Systems and methods for efficiently training a machine learning model are presented. More particularly, using information regarding the relevant neighborhoods of target nodes within a body of training data, the training data can be organized such that the initial state of the training data is relatively easy for a machine learning model to differentiate. Once trained on the initial training data, the training data is then updated such that differentiating between a matching and a non-matching node is more difficult. Indeed, by iteratively updating the difficulty of the training data and then training the machine learning model on the updated training data, the speed that the machine learning model reaches a desired level of accuracy is significantly improved, resulting in reduced time and effort in training the machine learning model.

Machine learning model training

Efficient processing of neighborhood data

Efficient generation of embedding vectors of nodes in a corpus graph

Systems and methods for generating embeddings for nodes of a corpus graph are presented. More particularly, embedding information of a target node may be based on the node itself, as well as related, relevant nodes to the target node within a corpus graph. The information of various nodes among the relevant nodes to the target node can be used to weight or influence the embedding information. Disclosed systems and methods include generating neighborhood embedding information for a target node, where the neighborhood embedding information includes embedding information from neighborhood nodes of the target node's relevant neighborhood, and where certain nodes having more relevance to the target node can be weighted to influence the generation of the neighborhood embedding information over nodes having less relevance to the target node.

Growing community of inventors

San Mateo, CA, United States of America

Kaifeng Chen