Clustering is a major field in machine learning and its aim is to divide the instances into different groups without any previous knowledge. The Density Peaks Clustering(DPC) has been presented in 2014. Although, DPC is effective, however, existing DPC based methods suffer from several issued such as: its sensitivity to the value user-adjustable parameter, using a simple kernel to calculate the local densities that can cause to choosing impropriate centers, and propagating errors which lead from the chain reaction. To tackle these issues in this paper a better method for selecting centers has been applied so that a set of cluster center candidates will be recognized from the dataset and the most proper one will be identified as main centers. Afterward, we generate a mutual kNN graph. The proposed graph is used to create a strong cluster backbone. We assign remaining candidates by computing their shortest path length to main centers then they will be assigned to which center that they have the least shortest path length, we propagate our labels by a graph-based label propagating method, the mutual kNN graph considers the data distribution, prevent form error propagation and can prevent noise points from taking any label.
