Density based spatial anomalous window discovery

The focus of this thesis is to identify anomalous spatial windows using clustering based methods. Spatial Anomalous windows are the contiguous groupings of spatial nodes which are unusual with respect to the rest of the data. Many scan statistics based approaches have been proposed for the identification of spatial anomalous windows. To identify similarly behaving groups of points, clustering techniques have been proposed. There are parallels between both types of approaches but these approaches have not been used interchangeably. Thus the focus of our work is to bridge this gap and identify anomalous spatial windows using clustering based methods. Specifically, we use the circular scan statistic based approach and DBSCAN to bridge the gap between clustering and scan statistics based approach. Our approach consists of the following steps: (a) Use the parameters proposed by DBSCAN to find core spatial nodes and its neighbors (b) Take combinations of nodes within a neighborhood to find smaller sub-sets of potentially anomalous windows (c) Take unions of all the combinations to explore bigger sub-sets of potentially anomalous windows. (d) Compute test-statistic for each of the window to identify its degree of unusualness. The window with the highest value of test statistic is the most unusual as compared to the rest of the data. We present extensive experimental results in US crime data set for various regions. Our results show that our approach is effective in identifying spatial anomalous windows and generally performs equal or better than existing scan statistic techniques and does better than a pure clustering method.