The original Affinity Propagation clustering algorithm (AP) only used the Euclidean distance of data sample as the only standard for similarity calculation. This method of calculation had great limitations for data with high dimension and sparsity when the original algorithm was running. Due to the single calculation method of similarity, the convergence and clustering accuracy of the algorithm were greatly affected. On the other hand, in the universe, we can consider the formation of galaxies is a clustering process. In addition, the interaction between different celestial bodies are achieved through universal gravitation. This paper introduced the Density Peak clustering algorithm (DP) and gravitational thought into the AP algorithm, and constructed the density property to calculate the similarity, put forward the Affinity Propagation clustering algorithm based on Gravity (GAP). The proposed algorithm was more accurate to calculate similarity of simple points through the local density of corresponding points, and then used the gravity formula to update the similarity matrix. The data clustering process could be seen as the sample points spontaneously attract each other based on ‘gravitation’. Experimental results showed that the convergence performance of GAP algorithm is obviously improved over the AP algorithm, and the clustering effect was better.
展开▼
