[4E3-GS-2] Machine learning: graph structure

[4E3-GS-2-01] Distributed Representation of Documents by Fusing Word Information and Link Information with Graph Autoencoders

〇Yoshiki Tanaka¹, Takashi Takekawa¹ (1. Kogakuin University of Technology & Engineering)

Keywords:distributed representation, Graph Autoencoder, graph

Converting web sites to vectors that include lots of information is useful. For instance, these vectors can be applied to recommender system. We obtained distributed representations of Wikipedia articles. Distributed representations of documents by conventional method have problem that there are cases where cosine similarity of unrelated articles is larger than that of related articles. In this work, we aimed to solve the problem by using information of links on web pages. A graph can represent the relationship among web pages. Therefore, using Graph Autoencoder and Variational Graph Autoencoder, we obtained distributed representations of documents, which are fused information of words and information of links. In these distributed representations, cosine similarities of related articles tend to be larger than those of unrelated articles. In addition, in vectors converted by Graph Autoencoders, those of articles belonging to same category tend to be closer than those by conventional methods.