Outlier Mining for Multivariate Time Series Based on Local Sparsity Coefficient

Multivariate time series (MTS) datasets are common in various financial, multimedia and medical applications. MTS samples which differ significantly from the remaining MTS samples are referred to as outliers. In this paper, the outlier mining method based on local sparsity coefficient for MTS is proposed. An extended Frobenius norm is used to compare the similarity between MTS samples, K-NN searches are performed by using two-phase sequential scan. MTS samples that are not possible outlier candidates are pruned, which reduces the number of computations and comparisons. The datasets of stock market and Australian Sign Language are used for outlier mining, the results show the effectiveness of the algorithm.