A decentralized gossip based approach for data clustering in peer-to-peer networks

In this paper, a novel distributed approach, named GDSOM-P2P, for clustering distributed data resources is proposed by combining, an improved version of Silhouette algorithm, the dynamic Self-Organizing Map (SOM) neural network, and VICINITY protocol as a generic overlay management framework based on self-organization. The proposed GDSOM-P2P is adapted to the dynamic conditions of these networks. In the proposed GDSOM-P2P algorithm, at first, each node extracts a number of important data through the SOM and Silhouette algorithms. Then each of the nodes chooses one of its neighbors with the help of the VICINITY algorithm, and exchanges their important data with their neighbors. By doing this, over a period, the nodes' data will be distributed over the entire network and the nodes in the network access the summary data model of the whole data. Finally, each node aggregates its internal data with a summary model and then performs the final clustering to cluster its internal data correctly. Evaluation results over a real P2P environment verify the efficiency of proposed GDSOM-P2P. Furthermore, the proposed GDSOM-P2P is also compared with the existing well-established distributed data clustering techniques. The results show a significant accuracy improvement of the proposed method.