Abstract: In order to get a low-cost subset of original features, we define the cost-distance among the samples and joint it to existing feature selection framework. We combine manifold learning into cost-sensitive feature selection model and develop a corresponding method, namely, cost-sensitive feature selection via manifold learning(CFSM). Most previous cost-sensitive feature selection algorithms rank features individually and select features just using correlation the between the cost and the features. Our cost-sensitive feature selection algorithm selects features not only using the correlation the between the cost and the features but also using the discriminative information implied within data to improve the features selection performance. Experimental results on different real world datasets show the promising performance of CFSM outperforms the state-of-the-arts.

Key words: cost-sensitive, manifold learning, feature selection, supervised learning