Abstract: Through the measurement of anomalies in multivariate spatio-temporal time sequences, it is possible to detect the anomalous regions from a large amount of data of the spatio-temporal events. Different from the techniques for detecting isolated anomalous data points, this paper proposes an unbiased KL divergence algorithm (UKLD). Firstly, the algorithm defines the divergent interval in the spatio-temporal time series. Gaussian distribution is used to estimate the distributions of the scanned interval and the remaining intervals after time-delay embedding, and the parameter estimation process of Gaussian distribution is sped up by using cumulative sums. Finally, the discrepancy level between intervals calculated by the unbiased KL divergence is used as the anomalous score of the scanned interval to obtain the spatio-temporal anomalous intervals. The simulation results show that, compared with HOT SAX algorithm and RKDE algorithm, UKLD is better for the spatio-temporal anomalous intervals detection task in terms of accuracy.

Key words: spatio-temporal data, anomalous regions detection, unbiased divergence, KL divergence