ShapeEvoNet: A shapelet-based Time Series Classification Method

Shapelet is the most representative subsequence in time series, and shapelet evolutionary or transitional patterns can provide semantic information for time series classification. In existing shapelet-based time series classification methods, shapelet extraction and evolutionary pattern construction are often separated, thus the evolutionary pattern which contains classification information cannot be used in the shapelet extraction stage. This paper proposes ShapeEvoNet, a time series classification method based on shapelets, which can extract shapelets and capture their evolutionary patterns simultaneously in an end-to-end model. First, we use Multi-length-input dilated causal Convolutional Neural Network (Mdc-CNN) to embed shapelet candidates of different lengths into a length-unified space, then we design a new cluster evolution triplet loss function to train the network in an unsupervised manner, which considers not only the external distance between anchor and positive (negative) samples, but also the internal distance within positive (negative) samples. We also construct a shapelet evolutionary graph among anchors and learn its transition probabilities automatically by Graph Attention Networks (GAT), so as to generate the shapelet embeddings with evolutionary information. Experimental results on 5 UCR public datasets and one real world dataset show that ShapeEvoNet outperforms existing state-of-the-art methods by 2.57% on accuracy.