Nearest neighbors-based adaptive density peaks clustering with optimized allocation strategy

Density peaks clustering (DPC) model is simple and effective in clustering data of any shape, and has attracted wide attention from scholars in recent years. However, it is difficult for DPC to determine the cutoff distance when calculating the local density of points, and to select the correct cluster centers of data with large differences of density between clusters or multi-density peaks in clusters; in addition, the point allocation method in DPC has low accuracy. To overcome these drawbacks, this paper presents a novel nearest neighbors-based adaptive DPC algorithm with an optimized allocation strategy (NADPC in short), and demonstrates its application in image clustering. First, the mutual nearest neighbor relationship between points is defined, the mutual neighborhood of point is proposed, and then a new local density of points is defined and does not need to set the cutoff distance. The candidate cluster centers and relative density are developed. According to the relative density and the high-density nearest neighbor distance of candidate cluster centers, their credibility as the cluster centers is calculated, and then the cluster centers are selected. Second, the mutual neighbor degree and similarity between two points are constructed. The neighborhoods of points are defined according to the high-density nearest neighbor, shared nearest neighbors, mutual neighbor degree and similarity, respectively. The similarity set, similarity domain, positive set, negative set, prediction set, positive value and predicted value of point are provided based on the above-mentioned neighborhoods. Then the optimized allocation strategy of points is proposed. Finally, the allocation algorithms of the non-abnormal and abnormal points are designed, respectively, and then the NADPC algorithm is designed. To evaluate the effectiveness of NADPC, it has been applied to 22 synthetic datasets and 26 actual datasets including 4 image datasets, and has great performance in terms of several evaluation metrics when compared with the other latest clustering algorithms.