Local dynamic neighborhood based outlier detection approach and its framework for large-scale datasets

Local outlier detection is a hot area and great challenge in data mining, especially for large-scale datasets. On the one hand, traditional algorithms often achieve low-quality detection results and are sensitive to neighborhood size. On the other hand, they are infeasible for large-scale datasets due to at least O(N2) time and space complexity. In light of these, we propose a new local outlier detection algorithm, which is designed based on a new stable neighborhood strategy-dynamic references nearest neighbors (DRNN). Meanwhile, we present a new detection framework by combining the proposed approach and k-mean for large-scale datasets. Experimental results demonstrate that the proposed algorithm can produce higher quality and robust detection results compared to several classic methods. Meanwhile, the new detection framework is able to significantly improve detecting efficiency without sacrificing accuracy.