Density Biased Sampling with Locality Sensitive Hashing for Outlier Detection

Outlier or anomaly detection is one of the major challenges in big data analytics since unusual but insightful patterns are often hidden in massive data sets such as sensing data and social networks. Sampling techniques have been a focus for outlier detection to address scalability on big data. The recent study has shown uniform random sampling with ensemble can boost outlier detection performance. However, uniform sampling assumes that all points are of equal importance, which usually fails to hold for outlier detection because some points are more sensitive to sampling than others. Thus, it is necessary and promising to utilise the density information of points to reflect their importance for sampling based detection. In this paper, we formally investigate density biased sampling for outlier detection, and propose a novel density biased sampling approach. To attain scalable density estimation, we use Locality Sensitive Hashing (LSH) for counting the nearest neighbours of a point. Extensive experiments on both synthetic and real-world data sets show that our approach significantly outperforms existing outlier detection methods based on uniform sampling.