[2S6-IS-3d] Machine learning

[2S6-IS-3d-02] A Study of Neighborhood Structure Consideration for Sequence Matching with Optimal Transport

〇Mitsuhiko Horie¹, Hiroyuki Kasai¹ (1. Waseda University)

Regular

Keywords:Sequence matching, Auto-weight learning, Optimal Transport

We propose a new distance measure between sequence data building upon optimal transport (OT) based sequence matching framework. The relationships between adjacent elements is an important feature for sequence data, and explicitly taking account of it is necessary. In addition to the costs for distances of data and its temporal order, the proposed distance defines the cost that considers the difference in the neighborhood structure of each element. Besides, the proposed distance automatically calculate the optimal weight of consideration of each costs for each pair of sequences. We conduct a classification experiment on some real-world datasets and show the effectiveness of the proposed distance compared to the state-of-the-art sequence matching methods.