Topologically-based Variational Autoencoder for Time Series Classification
Rodrigo Rivera · 2020
Topological Data Analysis is an approach to analyze data using different techniques from topology. These techniques aim to extract fundamental qualitative properties, such as shape and connectivity in data. In this work, we propose a universal approach for time series classification with variational autoencoders. It is built on extracted features from the persistent homology theory. Compared to standard classification approaches, the proposed methodology enables the classification of time series, which have different recurrent behavior in the reconstructed phase space. Multiple experiments with time-series datasets confirm that the method makes classification more robust to noisy and high-dimensional data and favors datasets in which shape has meaning.