Unsupervised feature extraction from multivariate time series for outlier detection

Kiyotaka Matsue, Mahito Sugiyama

Abstract: Although various feature extraction algorithms have been developed for time series data, it is still challenging to obtain a flat vector representation with incorporating both of time-wise and variable-wise association between multiple time series. Here we develop an algorithm, called Unsupervised Feature Extraction using Kernel and Stacking (UFEKS), that constructs feature vector representation for multiple time series in an unsupervised manner. UFEKS constructs a kernel matrix for the set of subsequences from each time series and horizontally concatenates all matrices. Then we can treat each row as a feature vector representation of its corresponding subsequence of times series. We examine the effectiveness of the extracted features under the unsupervised outlier detection scenario using synthetic and real-world datasets, and show its superiority compared to well-established baselines.