Improving ELM-Based Time Series Classification by Diversified Shapelets Selection

ELM is an efficient neural network which has extremely fast learning capacity and good generalization capability. However, ELM fails to measure up the task of time series classification because it hard to extract the features and characters of time series data. Especially, many time series has trend features which cannot be abstracted by ELM thus lead to accuracy decreasing. Although through selection good features can improve the interpretability and accuracy of ELM, canonical methods either fails to select the most representative and interpretative features, or determine the number of features parameterized. In this paper, we propose a novel method by selection diversified top-k shapelets to improve the interpretability and accuracy of ELM. There are three contributions of this paper: First, we put forward a trend feature symbolization method to extract the trend information of time series; Second, the trend feature symbolic expressions are mapped into a shapelet candidates set and a diversified top-k shapelets selection method, named as are proposed to find the most k distinguish shapelets; Last, we proposed an iterate ELM method, named as , automatically determining the best shapelets number and getting the optimum ELM classifier. The experimental results show that our proposed methods significantly improves the effectiveness and interpretability of ELM.