Abstract: Aiming at the problems that the truncation distance defined by the density peak clustering algorithm only considered the global distribution of samples and the "domino" phenomenon was easy to occur when assigning samples, a novel density peak clustering algorithm combining local truncation distance and small clusters merging was proposed. The truncation distance and local density of each sample were calculated based on the local distribution information of samples, which were conducive to accurately obtaining the density peaks on complex structure datasets. Potential density peaks were selected based on the difference between samples decision values and multiple small clusters were formed. A new kind of similarity between clusters was defined, and clusters were merged to obtain clustering results according to this similarity, which effectively avoided the "domino" phenomenon. Compared with several clustering algorithms on six synthetic datasets and eight UCI datasets, the standardized mutual information, adjusted rand index and adjusted mutual information average values of the proposed algorithm on 14 datasets were 18.15%, 28.99% and 20.22% higher than the five comparison algorithms on average, especially 30.06%, 47.15% and 31.90% higher than original density peak clustering algorithm. Experimental results showed the proposed algorithm had a good clustering effect.

Key words: clustering, density peak clustering, truncation distance, local density, potential density peaks