Granular-Ball Clustering Based Neighbourhood Outliers Detection Method

The detection of outliers and boundary points is a popular, valuable and interesting topic under any method, and often more important than normal points in some practical usage scenarios. Hyber Granular-Balls(GBs,This is what we will call it in the following) are a robust, scalable, adaptive class of spheres suitable for many frontier scenarios with applications ranging from astronomy to bioinformatics and pattern recognition. Currently, recognition methods based on correlation information from individual data are a common technique for outlier detection. In granular computing, efficiency and robustness to noise are proportional to the size of the granularity. Inspired by the idea of granularity computing, this paper uses GBs with multiple granularities instead of single data to obtain coarse-grained outlier features. Based on the structure and information of the GBs, we propose an adaptive coarse-grained outlier detection method: by detecting the number of points inside the GBs, the overlap, and the change in the neighbourhood relationship of each GB, we are able to produce effective detection identification of outliers. We have demonstrated the effectiveness of the GB method in several papers, and experimental results on synthetic and real datasets can also show that the method outperforms existing methods.