Explainable Framework for Time-series Analysis via Topological Data Analysis

Abstract

We propose an explainable framework for TDA-based time-series analysis, which characterizes time-series signals through time-delay embedding and persistent diagrams. Given the persistence diagram corresponding to a target class, our method continuously deforms an input signal into a signal whose diagram is close to the target diagram. We formulate this problem as a minimization of Wasserstein distance between persistence diagrams. The potential of this method is illustrated on some synthetic and real examples.

Tasks

Reproductions