A novel key-points based shapelets transform for time series classification

Time Series Classification (TSC) has been attracting more and more researchers attention. Shapelet is a novel concept proposed in recent years, which allows for TSC based on local features. A shapelet is subsequence of a time series and it could be maximally representative of a class. The original shapelet discovery algorithm is time-consume and the shapelet based classifier is constructed as a decision tree which non-leaf node contains the time-consume process. There are two shortcomings for the original algorithm: 1) shapelet discovery process is time-consume; 2) both feature extraction and classifier construction are tightly coupled, resulting in the novel concept is difficult to apply to construct other classifier. Although there are a mount of speed up technologies for the first shortcomings, we think that the newest proposed key-points based shapelet has the potential to resolve both poor interpretability and slow running problems. Whats more, the original algorithm remain a room for improvements. In this paper, we will propose two novelty. First, we will improve the original key-points based shapelet discovery algorithm. Second, we will apply the improved novel concept to shapelet transform algorithm. Besides the algorithm description, we also designed two groups of verified experiments. By comparing with state-of-the-art methods, we will demonstrate that our method could resolve the two shortcomings effectively.