A Shapelet Dictionary Learning Algorithm for Time Series Classification

Shapelet algorithm is a powerful time series classification method because it can extract discriminative subsequence from time series. Most existing Shapelet-based algorithm focus on searching Shapelet in training dataset, which brings two drawbacks about computationally expensive and low generalization ability. To solve the above drawbacks, this paper proposes a new algorithm named Shapelet Dictionary Learning (SDL), which can discover Shapelet in a generative way, not in a searching way like common Shapelet algorithm. The SDL algorithm optimizes the model of Dictionary Learning to generate a group of Shapelets, which has higher generalization ability. Compared with the baselines on 45 datasets, the proposed SDL algorithm achieves a significant improvement on the accuracy of time series classification.