Transfer Naive Bayes Learning using Augmentation and Stacking for SMS Spam Detection

Short Message Service (SMS) spam, unsolicited messages delivered through phones, is common and prevalent, but difficult to filter out. Naive Bayes (NB) classifier is a frequently used spam filtering approach for texts, due to its simple but rigorous statistical learning nature and transparency in the decision making. For SMS messages, simple NB classification is ineffective, because SMS texts are short and brief, often contain numerous typos, abbreviations, and slang words. In this paper, we propose, AstNB, a new Augmentation and Stacking combined Transfer learning approach for Naive Bayes (NB) classification. For effective transfer learning from a source domain, e.g. emails, to a target domain, e.g. SMS, AstNB first introduces data augmentation to generate different copies of training data, by combining a target domain sample with a randomly selected source domain instance, followed by training a number of basis classifiers from augmented data. After that, a stacking process is used to generate new feature space by aggregating predictions of basis classifiers and the feature space created from target data. A final classifier is trained to predict unlabeled SMS messages for spam prediction. Experiments and comparisons show that AstNB can effectively transfer knowledge from source domain for SMS spam detection, especially when the target domain has very few labeled messages.