Prediction of Category of Scientific Article by Graph Convolution

The convolution method that uses the IDs of citing paper and cited paper is known to improve the prediction performance of category of papers. This paper proposes a "word convolution" method that uses not only the IDs of the cited and citing papers, but also the words that appear in those papers. The proposed method improves the prediction performance (accuracy) 7% for the core dataset and 12% for the citeseer dataset and gives the same performance for the pubmed dataset compared with the state-of-the-art method.