[3G1-GS-2g] 機械学習：分類

[3G1-GS-2g-01] An Extended Model of Self-Attention Network That Enables Multi-label Node Classification

〇Reo Iizuka¹, Tatsuya Kawakami¹, Tianxiang Yang¹, Masayuki Goto¹ (1. Waseda University)

Keywords:Self-Attention Network, Multi-label Classification, GAT, Node Embedding, Graph Structure Data

In recent years, the development of information technology has led to the accumulation of data with graph structures, such as networks of papers. In graphs, the attributes of each node are assigned as labels, and the nodes often have multiple labels such as keywords of papers. Therefore, multi-label classification is important.
For instance, Heterogeneous Graph Attention Network(HAN) can classify a single node label, which uses a Self-Attention Network that takes account of the importance of neighboring nodes. However, it is difficult to obtain high accuracy by simply extending HAN to multi-label classification because some labels are not present in neighboring nodes.
In this study, we propose a multi-label classification method that uses the features obtained by SANNE, which captures a wide range of graph structures from its own nodes, as input for HAN. The effectiveness of the proposed method is demonstrated by applying it to the problem of predicting keywords of papers.