Density Peaks Clustering Based on Label Propagation and K-Mutual-Nearest Neighbors

The density peaks clustering algorithm is one of the density-based clustering algorithms. This algorithm has several advantages, including not requiring a preset number of clusters, requiring fewer parameters, and being able to achieve clustering of any shape. However, it also has limitations, such as poor clustering performance on datasets with uneven density, the need to manually select cluster centers on the decision graph, and a chain reaction that can lead to a large number of point misallocations due to incorrect allocation of individual points. To overcome the shortcomings of the density peaks clustering algorithm, we propose a density peaks clustering algorithm based on label propagation and k-mutual-nearest neighbors. First, the local density and the distance are defined by incorporating the concept of k-mutual-nearest neighbors to enhance clustering performance on datasets with uneven-density clusters. Second, an adaptive method for selecting cluster centers is proposed to avoid the manual selection of cluster centers. Third, an improved label propagation algorithm is used to assign all remaining points to solve the chain reaction problem. The experimental results show that our algorithm can accurately identify cluster centers and obtain high-quality clustering results on synthetic datasets with different characteristics, including datasets with uneven cluster density, convex datasets, manifold datasets, and datasets with inter-cluster contact. On different types of UCI datasets, including small datasets, high-dimensional datasets, and large datasets, our algorithm outperforms other comparative algorithms.