Classification in P2P Networks by Bagging Cascade RSVMs

Data mining tasks in P2P are bound by issues like scalability, peer dynamism, asynchronism, and data privacy preservation. These challenges pose difficulties for deploying conventional machine learning techniques in P2P networks, which may be hard to achieve classification accuracies comparable to regular centralized solutions. We recently investigated the classification problem in P2P networks and proposed a novel P2P classification approach by cascading Reduced Support Vector Machines (RSVM). Although promising results were obtained, the existing solution has some drawback of redundancy in both communication and computation. In this paper, we present a new approach to over the limitation of the previous approach. The new method can effectively reduce the redundancy and thus significantly improve the efficiency of communication and computation, meanwhile it still maintains good classification accuracies comparable to both the centralized solution and the previously proposed P2P solution. Experimental results demonstrate the feasibility and effectiveness of the new P2P classification solution.