Feature weighted confidence to incorporate prior knowledge into support vector machines for classification

Abstract

This paper proposes an approach called feature weighted confidence with support vector machine (FWC–SVM) to incorporate prior knowledge into SVM with sample confidence. First, we use prior features to express prior knowledge. Second, FWC–SVM is biased to assign larger weights for prior weights in the slope vector $\omega$ than weights corresponding to non-prior features. Third, FWC–SVM employs an adaptive paradigm to update sample confidence and feature weights iteratively. We conduct extensive experiments to compare FWC–SVM with the state-of-the-art methods including standard SVM, WSVM, and WMSVM on an English dataset as Reuters-21578 text collection and a Chinese dataset as TanCorpV1.0 text collection. Experimental results demonstrate that in case of non-noisy data, FWC–SVM outperforms other methods when the retaining level is not larger than 0.8. In case of noisy data, FWC–SVM can produce better performance than WSVM on Reuters-21578 dataset when the retaining level is larger than 0.4 and on TanCorpV1.0 dataset when the retaining level is larger than 0.5. We also discuss the strength and weakness of the proposed FWC–SVM approach.