Convolutional Nonlinear Neighbourhood Components Analysis for Time Series Classification

During last decade, tremendous efforts have been devoted to the research of time series classification. Indeed, many previous works suggested that the simple nearest-neighbor classification is effective and difficult to beat. However, we usually need to determine the distance metric (e.g., Euclidean distance and Dynamic Time Warping) for different domains, and current evidence shows that there is no distance metric that is best for all time series data. Thus, the choice of distance metric has to be done empirically, which is time expensive and not always effective. To automatically determine the distance metric, in this paper, we investigate the distance metric learning and propose a novel Convolutional Nonlinear Neighbourhood Components Analysis model for time series classification. Specifically, our model performs supervised learning to project original time series into a transformed space. When classifying, nearest neighbor classifier is then performed in this transformed space. Finally, comprehensive experimental results demonstrate that our model can improve the classification accuracy to some extent, which indicates that it can learn a good distance metric.