Publication - Patent

A Tag Inference System Based on Knowledge Graph Embedding

Tag Inference

Application No.

    Patent Pending

Authors

    Xueqi Cheng, Yantao Jia, Pengshan Cai, Manling Li, etc. (2nd student author)

Sponsored by

    Collaborative Research Program of Chinese Academy of Sciences and Huawei Inc. (one of the world's top 500 companies).

Abstract

  • Objective: Tag inference aims to predict tags based on the attributes of entities and their known tags.
  • Problem: Traditional tag inference methods are mainly based on Collaborative Filtering Recommendation, which can not formulate the semantic similarities of tags. Moreover, it handles "<entity - tag>" tuple and fails to handle "<entity - attribute - attribute value>" triples. Thus it will lead to performance degradation.
  • This patent proposed a tag inference method based on knowledge graph embedding, which jointly learns the embeddings of tuples and triples. Thus, it can capture semantic informantion and attribute informatin. Finally the tags are predicted by the score function based on the embedding vectors of entities and tags.

My Work

  • Constructed two subgraphs, i.e., "<entity - tag>" graph and "<entity - attribute - attribute value>" graph, and generated training data by selecting attributes that have similar distribution features with tags.
  • Speeded the model by employing the parallel framework ParTrans-X.
  • Applied the model successfully to the collaborative program of Chinese Academy of Sciences and Huawei Inc.