Representation Learning with Deconvolution for Multivariate Time Series Classification and Visualization.

We propose a new model based on the convolutional networks and SAX(Symbolic Aggregate Approximation) discretization to learn the representation for multivariate time series. The deep neural networks has excellent expressiveness, which is fully exploited by the convolutional networks with means of unsupervised learning. We design a network structure to obtain the cross-channel correlation with means of convolution and deconvolution, the pooling operation is utilized to perform the dimension reduction along each position of the channels. Discretization which based on the Symbolic Aggregate Approximation is applied on the feature vectors to extract the bag of features. We collect two different representations from the convolutional networks, the compression from bottle neck and the last convolutional layers. We show how these representations and bag of features can be useful for classification. We provide a full comparison with the sequence distance based approach on the standard datasets to demonstrate the effectiveness of our method. We further build the Markov matrix according to the discretized representation abstracted from the deconvolution, time series is visualized to complex networks through Markov matrix visualization, which show more class-specific statistical properties and clear structures with respect to different labels.