Abstract: This paper presented a dynamic version space division algorithm for SVM active learning, in view of the drawbacks of traditional batch sampling methods. Based on the duality property of feature space and parameter space, we discussed SVM active learning in dual space and concluded that example labeling in feature space corresponds to ver- sion space division in parameter space. Faking both the existing classification model and the previously labeled examples into consideration, we optimized the version space division process and maximized the value of selected examples for model refinement. In this way, a more effective selective sampling was achieved. Experimental results demonstrate the effectiveness of the dynamic version space division algorithm, which remarkably improves the classification performance under the cost of limited labeling effort.

Key words: Semi supervised learning, Active learning, Selective sampling, Support vector machine, Version space