Abstract:

An outlier detection algorithm based on Markov chain (MRKFOD algorithm) is presented. First, the basic data set is regarded as a weighted undirected graph, in which each datum represents a node, and each weighted edge denotes the similarity between nodes; so it forms an adjacency matrix, and then the adjacency matrix is regarded as a probability transition matrix in Markov chain. Secondly, the algorithm seeks the main feature vector of the probability transition matrix. Finally, the main feature vector of each node is looked upon as the outlier degree of each datum. The experimental results show that both the efficiency of MRKFOD algorithm and the maximum number of dimensions processed are obviously improved compared with other high-dimensional outlier mining algorithms.