Adaptive ensemble classification in p2p networks

Classification in P2P networks has become an important research problem in data mining due to the popularity of P2P computing environments. This is still an open difficult research problem due to a variety of challenges, such as non-i.i.d. data distribution, skewed or disjoint class distribution, scalability, peer dynamism and asynchronism. In this paper, we present a novel P2P Adaptive Classification Ensemble (PACE) framework to perform classification in P2P networks. Unlike regular ensemble classification approaches, our new framework adapts to the test data distribution and dynamically adjusts the voting scheme by combining a subset of classifiers/peers according to the test data example. In our approach, we implement the proposed PACE solution together with the state-of-the-art linear SVM as the base classifier for scalable P2P classification. Extensive empirical studies show that the proposed PACE method is both efficient and effective in improving classification performance over regular methods under various adverse conditions.