Deep Temporal Clustering: Fully unsupervised learning of time-domain features

Unsupervised learning of timeseries data is a challenging problem in machine learning. Here, we propose a novel algorithm, Deep Temporal Clustering (DTC), a fully unsupervised method, to naturally integrate dimensionality reduction and temporal clustering into a single end to end learning framework. The algorithm starts with an initial cluster estimates using an autoencoder for dimensionality reduction and a novel temporal clustering layer for cluster assignment. Then it jointly optimizes the clustering objective and the dimensionality reduction objective. Based on requirement and application, the temporal clustering layer can be customized with any temporal similarity metric. Several similarity metrics are considered and compared. To gain insight into features that the network has learned for its clustering, we apply a visualization method that generates a heat map of regions of interest in the timeseries. The viability of the algorithm is demonstrated using timeseries data from diverse domains, ranging from earthquakes to sensor data from spacecraft. In each case, we show that our algorithm outperforms traditional methods. This performance is attributed to fully integrated temporal dimensionality reduction and clustering criterion.