Learnable Dynamic Temporal Pooling for Time Series Classification

With the increase of available time series data, predicting their class labels has been one most important challenges in a wide range disciplines. Recent studies on classification show that convolutional neural networks (CNN) achieved state-of-the-art performance as single classifier. In this work, pointing out global pooling layer is usually adopted by existing CNN classifiers discards temporal information high-level features, we present dynamic (DTP) technique reduces size hidden representations aggregating features at segment-level. For partition whole into multiple segments, utilize warping (DTW) to align each point order with prototypical which can be optimized simultaneously network parameters classifiers. The DTP combined fully-connected helps extract further discriminative considering position within an input series. Extensive experiments both univariate and multivariate datasets our proposed significantly improves performance.