A Proximity Forest for Multivariate Time Series Classification

Multivariate time series (MTS) classification has gained attention in recent years with the increase of multiple temporal datasets from various domains, such as human activity recognition, medical diagnosis, etc. The research on MTS is still insufficient and poses two challenges. First, discriminative features may exist on the interactions among dimensions rather than individual series. Second, the high dimensionality exponentially increases computational complexity. For that, we propose a novel proximity forest for MTS classification (MTSPF). MTSPF builds an ensemble of proximity trees that are split through the proximity between unclassified time series and its exemplar one. The proximity of trees is measured by locally slope-based dynamic time warping (DTW), which enhances traditional DTW by considering regional slope information. To extract the interaction among dimensions, several dimensions of an MTS instance are randomly selected and converted into interrelated sequences as the input of trees. After constructing each tree independently, the weight of each tree is calculated for weighted classifying. Experimental results on the UEA MTS datasets demonstrate the efficiency and accuracy of the proposed approach.