Time-Series Classification with Constrained DTW Distance and Inverse-Square Weighted k-NN

The problem of time-series classification witnessed the application of many techniques for data mining and machine learning, including neural networks, support vector machines, and Bayesian approaches. Somewhat surprisingly, the simple 1-nearest neighbor (1NN) classifier, in combination with the Dynamic Time Warping (DTW) distance measure, is still competitive and not rarely superior to more advanced classification methods, which includes the majority-voting k-nearest neighbor (kNN) classifier. In this paper we focus on the kNN classifier combined with the inverse-squared weighting scheme, and its interaction with constrained DTW distance. By performing experiments on the entire UCR Time Series Classification Archive we show that with proper selection of the constraint parameter $r$ and neighborhood size $k$ , inverse-square weighted kNN consistently outperforms 1NN.