Readable contrast mining method for heterogeneous bipartite networks with class label

Recent IoT trends have been driving usage of multi-source data. In general, since the multi-source data have different nature, we need to make heterogeneous networks to represent them simultaneously. This paper presents a novel mining method to discover factors that characterize differences between classes for heterogeneous bipartite networks. In order to find the differences from such the large-scale networks efficiently, it is important to use distributed representations that preserve first-order and second-order proximity between vertices of the networks. We propose an effective representation method for heterogeneous bipartite networks with class label. And we propose a readable visualization method for revealing the factors on the embedding space. From computational experiments using real-world data which include the multimedia access logs and the results of questionnaire for the users, we show that the proposed method can acquire distributed representations with higher explanatory power than existing methods, and can discover important factors.