Detecting Unusual Pattern with Labeled Data in Two-Stage

More analysis has been done to discover the meaningful unusual patterns which may mean fraud or anomaly. In this paper, a two-stage approach considering the labeled data´s proposed to discover meaningful unusual observation, without the goal of classifying. We firstly apply hyper surface classification (HSC) algorithm to gain a separating hyper surface which includes several pieces. Observation in the sparse piece is viewed as the unusual pattern. For therpieces with local density, we construct a weighted graph for it and search the minimum spanning tree (MST), then detect further by cutting off several edges with the aximum weight. Combining the advantages of the two stages, a process of subdividing is proposed to consider the domain knowledge. Experimental results show that our approach can detect unusual pattern effectively together with other hidden valuable knowledge.