Support vector machine learning from positive and unlabeled samples

In many machine learning settings, labeled samples are difficult to collect while unlabeled samples are abundant. We investigate in this paper the design of support vector machine classification algorithms learning from positive and unlabeled samples only. We first find the minimum bounding sphere that enclosed all the positive samples, and then use this minimum bounding sphere to pick out the negative samples from the unlabeled samples, at last we train the support vector machine using the training set which consists of the given positive samples and the negative samples picked out from the unlabeled samples. Experiments indicate that support vector machine learning from positive and unlabeled samples achieves the desired high test precision and prediction accuracy.