A Graph-based Density Peaks Method by Employing Shortest Path for Data Clustering

摘要

Data clustering is one of the most important and fundamental tasks of machine learning. Data clustering aims at dividing a set of objects into several groups according to their similarities. In recent years Density Peaks Clustering (DPC) was introduced as a fast and non-iterative clustering method which does not require any previous knowledge about the number of clusters. However, this method suffers from a few shortcomings such as its sensitivity to the user-adjustable parameter, disability to consider data distribution, and inappropriate center selection when facing complex clusters. To overcome these issues, in this paper, a novel density-based peaks clustering method called GDPCS is proposed. By employing the properties of the mutual neighborhood graph and shortest path distance, the proposed method considers the data distribution, present a better shape of clusters, and reduces the clusters' connectivity. To demonstrate the proposed method's effectiveness and superiority, many experiments were performed on both real-world and synthetic datasets. The obtained results show that the proposed method has achieved an acceptable result on imbalanced and complex shaped clusters and can detect more appropriate centers.