Integrating Data-Driven Segmentation, Local Feature Extraction and Fisher Kernel Encoding to Improve Time Series Classification

The uniform sampling strategy is widely used in time series segmentation, but unable to handle time warping problem or preserve the latent patterns in time series. To solve these shortcomings, a brand new data-driven segmentation method is proposed, which could segment time series into subsequences with different lengths adaptively. Then a time series classification method under the bag-of-word framework is proposed. Two kinds of mutually complementary features, i.e., interval feature and normal cloud model feature, are extracted from subsequences. And then time series are encoded into Fisher Vectors. Finally, a linear support vector machine is used as the classifier. Experiments on 43 UCR datasets show that the newly proposed method has promising classification accuracies comparing with state of the art baselines. Moreover, due to the data-driven segmentation and timesaving local feature extraction, the method has low time complexity, which is also demonstrated in the experiments.