Dual Increment Shapelets: A Scalable Shapelet Discovery for Time Series Classification

Abstract

As time series data become more complex and users expect more sophisticated information, numerous algorithms have been proposed to solve these challenges. Among those algorithms to classify time series data, shapelet – a discriminative subsequence of time series data – is considered a practical approach due to its accurate and insightful classification. However, previously proposed shapelet algorithms still suffer from exceedingly high computational complexity, as a result, limiting its scalability to larger datasets. Therefore, in this work, we propose a novel algorithm that speeds up shapelet discovery process. Our algorithm so called “Dual Increment Shapelets (DIS)” is a combination of two-layered incremental neural network and filtering process based on subsequence characteristics. Empirical experiments on forty datasets evidently demonstrate that our proposed work could achieve large speedup while maintaining its accuracy. Unlike the previous algorithm that mainly emphasizes speedup of the search algorithm, DIS essentially reduces the number of shapelet candidates based on subsequence characteristics. As a result, our DIS algorithm could achieve more than three orders of magnitude speedup, comparing with the baseline algorithms, while preserving the accuracy of the state-of-the-art algorithm.