Extracting Distinctive Shapelets with Random Selection for Early Classification

Abstract

Early classification on time series has attracted much attention in time-sensitive domains. The goal of early classification on time series is to achieve better classification accuracy, and meanwhile to make prediction as early as possible. Shapelets are local features of time series and have high discriminability. In shapelet-based early classification, due to the large number of shapelet candidates, it is challenging to discover shapelets more effectively. In this paper, we propose Early Random Selection Shapelet Classification on Time Series (EARSC). Firstly, we identify the representative time series for each class. Secondly, we extract shapelet candidates for the representative time series and then evaluate them to obtain prior knowledge. Thirdly, we design random selection strategy with prior knowledge to select the better shapelet and make early classification. Experimental results on 14 real datasets have shown the effectiveness of the proposed method.