Boundary Point Detection Combining Gravity and Outlier Detection Methods

Vijdan Khalique, Hiroyuki Kitagawa, Toshiyuki Amagasa

Abstract: Boundary point detection is the task of identifying points occurring at the boundary of a dense region in a dataset. It can reveal useful information about the system generating the data. Several data mining methods have been proposed to solve this problem. In our previous work, we proposed the Boundary Point Factor (BPF), which combined Gravity and an outlier detection method; Local Outlier Factor (LOF) to calculate the BPF score to detect the boundary points. The method is effective in a variety of real and synthetic datasets. However, one of the most crucial questions is whether other outlier detection methods can be used with Gravity for boundary point detection. In this work, we first investigate the favorable properties of LOF that make it suitable to be combined with Gravity for detecting boundary points. Next, the comparison of the commonly used outlier detection methods with the useful properties of LOF is performed on various datasets to demonstrate if these methods can be used for boundary point detection. Overall, it was found that, unlike LOF, a straightforward combination of other outlier detection methods with Gravity cannot be used, and sophisticated manipulation of the outlier scores generated by these methods may be needed to enable them to be used with Gravity for boundary point detection.