[3G1-GS-2g] 機械学習：分類

[3G1-GS-2g-02] A Study on the Construction Method of Shared Structure for Deep Neural Network in Multi-Label Classification

〇Kodai Ishikura¹, Aya Kitasato¹, Gendo Kumoi¹, Masayuki Goto¹ (1. Waseda university)

Keywords:Machine Learning, Deep Neural Learning, Multi Lable Classification, DNN Structure, Clustering

Recently, the importance of techniques related to multi-label classification, which assumes that multiple labels are assigned to a single document, has been increasing. One of the approaches to solve this problem is Branched Multi-Task Networks (BMTN), which constructs a network in which the middle layer of the Deep Neural Network is shared by labels that are highly related. In BMTN, the shared structure is determined by clustering the similarity between the labels, but the number of clusters in each middle layer must be set in advance by the analyst. Therefore, it doesn't adequately represent the relationship between labels. In this study, we propose an algorithm for determining the number of clusters that can adequately represent the relationship between labels in clustering. Finally, we apply the proposed method to the Yomiuri article data, and show the usefulness of the proposed method in terms of estimation accuracy.