Multi-Scale Change Point Detection in Multivariate Time Series

A core problem in time series data is learning when things change. This problem is especially challenging when changes appear gradually and at varying timescales, such as in health. Convolutional Neural Networks (CNNs) have the potential to recognize and localize complex patterns, but are sensitive to scale. We propose a new class of scale and shift invariant neural networks that augment CNNs with trainable wavelet layers. Experimentally, we demonstrate that this approach can be used to more accurately detect gradual change points in multivariate time series.