Deep semi-supervised clustering for multi-variate time-series

Huge amount of data are nowadays produced by a large and disparate family sensors, which typically measure multiple variables over time. Such rich information can be profitably organized as multivariate time-series. Collect enough labelled samples to set up supervised analysis for such kind is challenging while reasonable assumption dispose limited background knowledge that injected in the process. In this context, semi-supervised clustering methods represent well suited tool get most out reduced knowledge. With aim deal with time-series under setting, we propose (constrained) deep embedding framework exploits supervision modeled Must- Cannot-link constraints. More detail, our proposal, named conDetSEC (constrained Deep time SEries Clustering), based on Gated Recurrent Units (GRUs) explicitly manage temporal dimension associated multi-variate series data. implements procedure an generation step combined refinement step. Both steps exploit small available provided specifically, during constraints used jointly optimizing network parameters via both unsupervised tasks, at they conjunction goal stretch manifold towards centroids recover more clear cluster structure. Experimental evaluation real-world benchmarks coming from diverse domains has highlighted effectiveness proposal comparison state-of-the-art methods.