A density-core-based clustering algorithm with local resultant force

Clustering analysis has been widely used in image segmentation, face recognition, protein identification, intrusion detection, document clustering and so on. Most of the previous clustering algorithms are not suitable for complex situations with manifold structure and large variations in density. Clustering by density core (DCore) turns out to be a very effective clustering method for complex structure. However, DCore must set too many parameters for better results, which often fails when the shape of data is complex and the density of data varies too much. Inspired by universal gravitation, we propose a novel clustering algorithm (called DCLRF) based on density core and local resultant force. In this algorithm, each data point is viewed as an object with a local resultant force (LRF) generated by its neighbors and a local measure named centrality is proposed based on LRF and natural neighbors. Firstly, we extract core points using the CE value. Then, we use the natural neighbor structure information of core points to get the final clustering results. Excluding the influence of noise, core points can well represent the structure of clusters. Therefore, DCLRF can obtain the optimal cluster numbers for the datasets which contain clusters of arbitrary shapes. Both synthetic datasets and real datasets are used for experiments to verify the efficiency and accuracy of the DCLRF.