For Bulletin Board System(BBS) sentiment classification issues, an improved Random Subspace Method(RSM) is proposed. This method tries to make full use of the discriminative information in the high dimensional feature space. In the process of generating subspaces, on the one hand, a weighting function is used to evaluate classification abilities of the features, and better ones are chosen to ensure accuracy of classification with a higher probability, on the other hand, the size of the subspace is enlarged, principal component analysis is u...