An Effective Lazy Shapelet Discovery Algorithm for Time Series Classification

Wei Zhang, Zhihai Wang, Jidong Yuan, Shilei Hao

Abstract: Shapelet is a primitive for time series classification. As a discriminative local characteristic, it has been studied widely. However, global shapelet-based models have some obvious drawbacks. First, the progress of shapelet extraction is time consuming. Second, the shapelets discovered are merely good on average for the training instances, while local features of each instance to be classified are neglected. For that, instance selection strategy is used to improve the efficiency of feature discovery, and a lazy model based on the local characteristics of each test instance is proposed. Different from the commonly used nearest neighbor models based on global similarity, our model alleviates the uncertainty of predicted class value using local similarity. Experimental results demonstrate that the proposed model is competitive to the benchmarks and can be effectively used to discover characteristics of each time series.