Machine Learning Assisted Temporal Continuous Clustering

Unsupervised learning of temporal sequences is an essential data mining technique to explore and find groups of similar time series objects. The application appears in numerous fields of science like in finance, engineering, healthcare, etc. Traditionally feature-based clustering required domain knowledge to extract high quality features. Now deep autoencoder models can achieve dimension reduction and learn good representations in an unsupervised fashion at the same time. These models with a well-chosen architecture can be applicable for many different data types, eliminating the strong reliance on domain specific knowledge for feature extraction that later used for clustering purpose. This paper presents the evaluation of the Deep Continuous Clustering method for processing continuous data series. As visible from the results, it has significantly better accuracy than the widely used DBSCAN method.