Highly Scalable SVM Modeling with Random Granulation for Spam Sender Detection

Spam sender detection based on email subject data is a complex large-scale text mining task. The dataset consists of email subject lines and the corresponding IP address of the email sender. A fast and accurate classifier is desirable in such an application. In this research, a highly scalable SVM modeling method, named Granular SVM with Random granulation (GSVM-RAND), is designed. GSVM-RAND applies bootstrapping to extract a number of subsets of samples from the original training dataset. Each training subset is then projected into a feature subspace randomly selected from the original feature space. Here we call a granule such a subset of samples in such a feature subspace. A local SVM is then modeled in each granule. For a new sample, it is firstly projected into each granule in which the local SVM is fired to make a prediction. After that, all SVM predictions are aggregated by Bayesian Sum Rule for a final decision. GSVM-RAND is easy to be parallelized and hence efficient and highly scalable. GSVM-RAND is also effective by integrating a large number of weak, low-correlated local SVMs.