KAGO: an approximate adaptive grid-based outlier detection approach using kernel density estimate

Outlier detection approaches show their efficacy while extracting unforeseen knowledge in domains such as intrusion detection, e-commerce, and fraudulent transactions. A prominent method like the K-Nearest Neighbor (KNN)-based outlier detection (KNNOD) technique relies on distance measures to extract the anomalies from the dataset. However, KNNOD is ill-equipped to deal with dynamic data environment efficiently due to its quadratic time complexity and sensitivity to changes in the dataset. As a result, any form of redundant computation due to frequent updates may lead to inefficiency while detecting outliers. In order to address these challenges, we propose an approximate a daptive g rid-based o utlier detection technique by finding point density using k ernel density estimate (KAGO) instead of any distance measure. The proposed technique prunes the inlier grids and filters the candidate grids with local outliers upon a new point insertion. The grids containing potential outliers are aggregated to converge on to at most top-N global outliers incrementally. Experimental evaluation showed that KAGO outperformed KNNOD by more than an order of ≈ 3.9 across large relevant datasets at about half the memory consumption.