A local-gravitation-based method for the detection of outliers and boundary points

Detection of outliers and boundary points represents an effective, interesting and potentially valuable pattern, which may be more important than that of normal points. In order to detect outliers and boundary points, we propose a local-gravitation-based method in which each data point is viewed as an object with both mass and a local resultant force (LRF) generated by its neighbors. With the increase of neighbor, the LRF of outliers, boundary points and interior points varies at different rates. In this paper, the LRF changing rates of points with lower densities have higher scores, namely the changing rate of an outlier is greater than that of a boundary point and inner point. In other words, top-m ranked points can be identified as outliers, and the greater the LRF changing rate of a point is, the more likely it is a boundary point. The main advantage of our proposed method is that it does not depend on the choice of K value, which improves the detection performance. The experimental results on synthetic and real data sets show that the proposed method is better than the existing methods.