Simultaneous Visualization of Documents, Words and Topics by Tensor Self-Organizing Map and Non-negative Matrix Factorization

The purpose of this work is to develop a simultaneous visualization method of documents, words, and topics. The task of the proposed method is to map a set of documents to a pair of low-dimensional latent spaces corresponding to documents and words, by which the relations between them are visualized. In addition, the method also decomposes the mapping as the sum of topics, so that the topic distributions are visualized on the latent spaces. To achieve the task, we combined the tensor self-organizing map and the non-negative matrix factorization. We applied the method to NeurIPS data set, and the result shows that the method enables us to understand the tripartite relation between document, words and topics easily.