UNSUPERVISED TIME-SERIES CLASSIFICATION

In this paper a scheme for unsupervised probabilistic time series clas sification is detailed. The technique utilizes autocorrelation terms a s discriminatory features and employs the Volterra Connectionist Model (VCM) to transform the multi-dimensional feature information of each t raining vector to a one-dimensional classification space. This allows the probability density functions (PDFs) of the scalar classification indices to be represened as a function of the classifier weights. The weight values are chosen so as to maximize the separability of the cla ss conditional PDFs. Statistical similarity tests based on the overlap area of the PDFs are then performed to determine the class membership of each training vector. Results are presented that illustrate the pe rformance of the scheme applied to synthetic and real world data.