Incremental Local Evolutionary Outlier Detection for Dynamic Social Networks

Numerous applications in dynamic social networks, ranging from telecommunications to financial transactions, create evolving datasets. Detecting outliers in such dynamic networks is inherently challenging, because the arbitrary linkage structure with massive information is changing over time. Little research has been done on detecting outliers for dynamic social networks, even then, they represent networks as un-weighted graphs and identify outliers from a relatively global perspective. Thus, existing approaches fail to identify the objects with abnormal evolutionary behavior only with respect to their local neighborhood. We define such objects as local evolutionary outliers, LEOutliers. This paper proposes a novel incremental algorithm IcLEOD to detect LEOutliers in weighted graphs. By focusing only on the time-varying components (e.g., node, edge and edge weight), IcLEOD algorithm is highly efficient in large and gradually evolving networks. Experimental results on both real and synthetic datasets illustrate that our approach of finding local evolutionary outliers can be practical.