Real-world Data Clustering Based on Dominant Set and Nearest Neighbors

Chongwei Huang, Jian Hou

Abstract: In most clustering algorithms, determining the boundary extension of clusters is a common challenge and most of methods show undesirable outcome. Thus, our algorithm combines the centroid drift technique with natural nearest neighbor to extract the boundary of cluster, ensuring that the data within each boundary is independent of external data. Next, we use dominant set clustering to extract the core data of each cluster and obtain small adaptive weight thresholds from the dominant sets to include more data points. Finally, we allocate external elements to the subclusters with the highest membership. To validate the effectiveness of our algorithm, we conducted experiments on 9 real datasets. Furthermore, we compare our algorithm with currently popular algorithms. It demonstrates promising results of our approach.