Avoiding the Cluster Hypothesis in SV Classification of Partially Labeled Data

We propose a Support Vector-based methodology for learning classifiers from partially labeled data. Its novelty stands in a formulation not based on the cluster hypothesis, stating that learning algorithms should search among classifiers whose decision surface is far from the unlabeled points. On the contrary, we assume such points as specimens of uncertain labels which should lay in a region containing the decision surface. The proposed approach is tested against synthetic data sets and subsequently applied to well-known benchmarks, attaining better or at least comparable performance w.r.t. methods described in the literature.