Time Series Data Classification Using Discriminative Interpolation with Sparsity

Here we consider a novel approach for the categorization of time series data, called Classification by Discriminative Interpolation with Sparsity (CDIS), that circumvents the need for feature extraction as in traditional machine learning techniques. During training, the wavelet representations of functions in the same class are warped to become more similar to each other while moving away from functions in different classes. This process—termed discriminative interpolation—leads to gerrymandering of functional neighborhoods in service of supervised learning. We detail a full multiresolution wavelet expansion, incorporated with sparsity, for the functional data. The utility of the proposed CDIS method was experimentally validated on several data sets, thus demonstrating its competitiveness against contemporary and state-of-the-art feature-based methods.