A Novel Clustering Algorithm based on Directional Propagation of Cluster Labels

Clustering is an important topic in the field of machine learning. There are abundant algorithms which are proposed for clustering, and they are mainly based on two basic physical metrics, distance and density. However, it is difficult to reflect the orientation relationship between data by distance and density alone, while the orientation relationship can be easily expressed by direction. Inspired by this, we regard direction as the main basic physical metric, and propose a clustering algorithm based on directional propagation of cluster labels, namely DBC (Direction Based Clustering). DBC doesn't need to make specific assumptions about the density of data, but uses the orientation relationship between data to help clustering. Similar to the density-based clustering algorithms DBSCAN (Density Based Spatial Clustering Of Applications with Noise) and DPC (Density Peak Clustering), DBC can also find clusters of arbitrary shapes. It requires two parameters, but its two parameters have a fixed range of empirical values. Moreover, it is less affected by uneven density distribution than DBSCAN and DPC. We compare DBC with four well-known clustering algorithms, including DPC, DBSCAN, AP (Affinity Propagation) and K-means++. Experiments on artificial data sets show that DBC performs better than DBSCAN and DPC on data sets with uneven density distribution, and can effectively identify clusters with overlapping regions. Experiments on real-world data sets show that DBC has advantages over DPC, AP and K-means++ in clustering effect.