A Novel Random Subspace Method for Online Writeprint Identification

With the widespread application of computer network technology, diverse anonymous cyber crimes begin to appear in the online community. The anonymous nature of online-information distribution makes writeprint identification a critical forensic problem. But the difficulty of the task is the huge number of features in even a moderate-sized available text corpus, which causes the problem of over-training. In this paper, we proposed a novel random subspace method by constructing a set of stable classifiers to take advantage of nearly all the discriminative information in the high dimensional feature space. In the construction of base classifiers, an optimized synergetic neural network is employed to provide probabilistic information for each class. Performance results on the subset of Reuters Corpus Volume 1 (RCV1) show that the proposed random subspace method achieves the better identification performance than a single classifier and conventional random subspace methods.