A new adaptive and effective granular ball generation method for classification

Granular ball computing is an effective, robust and scalable multi-granularity learning method, which consists of two phases: granular ball (GB) generation and GB-based learning. However, the majority of existing GB generation methods suffer from the limitations of being non-adaptive or generating GBs of inferior quality. Moreover, GB-based classification rule needs to be further optimized as it ignores the possible overlap between GBs. In this paper, a new adaptive and effective GB generation method based on the shortest heterogeneous distance (ADPGBG) is proposed. It can effectively improve the quality of GBs and the GB generation process is completely parameter-free. Subsequently, a corresponding outlier detection method is introduced to weaken the impact of noisy data during the GB generation process. Furthermore, an improved GB k-nearest neighbors classifier (IGBkNN) is developed based on the ADPGBG method, which refines the classification rule and addresses the issue that the queried samples are difficult to be classified correctly in the overlapping region between heterogeneous GBs by introducing the concept of density of the GB. Finally, to evaluate the performance of IGBkNN, extensive experiments are performed and the experimental results demonstrate the superiority of the proposed method compared to the mainstream GB generation methods for classification.