Unsupervised feature extraction from multivariate time series for outlier detection

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 series. Here we develop an algorithm, called Unsupervised Feature Extraction using Kernel Stacking (UFEKS), that constructs in unsupervised manner. UFEKS kernel matrix the set subsequences from each horizontally concatenates all matrices. Then can treat row as its corresponding subsequence times We examine effectiveness extracted features under outlier detection scenario synthetic real-world datasets, show superiority compared well-established baselines.