Adaptive Density Peaks Clustering Based on Relative-Distance Growth Rate and Density-Bridge Merging

Abstract

Density peaks clustering (DPC) is a representative density-based clustering method. However, conventional DPC usually relies on manual cluster-center selection or prior knowledge of the cluster number, and may suffer from unstable label assignment on complex data distributions. To improve the automation capability and structural adaptability of DPC, this paper proposes an adaptive density peaks clustering algorithm based on relative-distance growth rate and density-bridge merging, termed ADPC-GB. In the first stage, the algorithm first constructs a local reference relative distance from the neighborhood of the nearest higher-density point. It then compares the current relative distance with this local reference and calculates the relative-distance growth rate. Samples whose growth rates meet or exceed the predefined threshold are identified as initial cluster centers. In the second stage, density-bridge strength and the local-scale distance ratio are introduced to evaluate density continuity and local geometric consistency between adjacent initial clusters, respectively. Based on these two criteria, over-segmented subclusters belonging to the same underlying structure can be merged, while false merging caused by accidental proximity or weak density connections can be reduced. Experimental results on synthetic and real-world datasets show that ADPC-GB can identify intrinsic cluster structures without manually specifying the number of clusters, and achieves competitive clustering accuracy and stability on complex non-convex, density-imbalanced, and boundary-adjacent datasets.