A Proposal for Shape Aware Feature Extraction for Time Series Classification

Many new classification methods are proposed for time series data in this decade, including ensembles and deep learning based methods. However, speed for real-time classification, interpretability of the results and necessary computational resources, make them difficult to use in real life problems compared to traditional feature based or similarity based time series classification methods. Judicious use of local features of time series is supposed to be the key point to improve the performance and interpretability. In this work, three new linear time complexity shape aware feature extraction methods leading to the computation of similarities of two time series, are proposed. Their performances are compared to the most popular Dynamic Time Warping (DTW) with the k- Nearest Neighbor classifier (kNN) as the baseline classifier (kNN-DTW) by simulation experiments with 43 benchmark time series data sets. It is found that the proposed approach can achieve higher classification accuracies for some datasets while computationally lighter for all the data sets than DTW.