Detecting anomalies in symbolic sequence dataset

This paper presents an anomaly detection technique for symbolic sequences based on support vector data description (SVDD). It introduces the longest common subsequence (LCS) distance metric to compute pairwise similarity between sequences. We extend the kernel method to the analysis of variable-length sequences by embedding the LCS distance into the form of Gaussian function to develop a novel Gaussian LCS kernel. By using this kernel, SVDD can directly handle input sequences of variable length, and make good use of the sequential information of sequences. The performance of the proposed technique was compared with SVDD with Gaussian RBF kernel and spectrum kernel. Experimental results show that this technique is better than other techniques in achieving higher detection rate and lower false positive rate.