Title:Topological feature selection for time series data

Abstract:We use tools from applied topology for feature selection on time series data. Given a multivariate time series, we employ persistent homology to assign weights to individual components and construct weighted embeddings whose topological features are optimized according to a chosen objective. We show that, along affine paths in the space of weights, the persistence diagram evolves piecewise linearly. This enables explicit and efficient computation of directional derivatives and gradient-based optimization. We prove that for objective functions on persistence diagrams that are Lipschitz with respect to a $p$-Wasserstein distance, the evaluation of the function along a gradient-descent path is Lipschitz. We apply our method to synthetic and biological data sets and demonstrate that our approach selects informative combinations of variables that recover the coordinated dynamics.