A DBN-Based Classifying Approach to Discover the Internet Water Army

The Internet water army (IWA) usually refers to hidden paid posters and collusive spammers, which has already generated big threats for cyber security. Many researchers begin to study how to effectively identify the IWA. Currently, most efforts to distinguish non-IWA and IWA in data mining context focus on utilizing classification-based algorithms, including Bayesian Network, SVM, KNN and etc... However, Bayesian Network need strong conditional independence assumption, KNN has big computation costs, above approach may affect the effectiveness to some extent in real industrial applications. Hence, Neural Networks-like deep approach for IWA identification gradually becomes an emerging but possible direction and attempt. Unfortunately, there also exists one main problem, which is how to balance the deep learning and computation costs in hierarchical architecture. More specially, combine leaning-level heuristic training design and computing-level concurrent computation is a challenging issue. In this paper, we propose a collaborative hierarchical approach based on the deep belief network (DBN) for IWA identification. Firstly, a DBN-based collaborative model with hierarchical classifying mechanism is built. Then towards Hadoop platform, the Downpour Stochastic gradient descent (Downpour SGD) is exploited for DBN pre-training. Finally, the dynamical workflow will be designed for managing the whole learning-based classifying process. The experimental evaluation shows that the valid of our approach.