Single-class classification augmented with unlabeled data: A symbolic approach

Abstract

Supervised machine learning techniques generally require that the training set on which learning is based contain sufficient examples representative of the target concept, as well as known counterexamples of the concept; however, in many application domains it is not possible to supply a set of labeled counter-examples. This could be due to either the nature of the problem domain, or the expense of obtaining labels for the training data. This paper presents a technique that can be used for learning symbolic concept descriptions from a dataset consisting of labeled positive examples together with a corpus of unlabeled positive and negative examples. The technique uses evolutionary search to explore the space of concept descriptions, guided by an evaluation function that seeks to achieve a balance between the generalization and specialization capacities of individuals. One of the features of the technique is that it requires empirical determination of a bias weighting factor that controls the relative balance between the generalization and specialization of hypotheses. The advantage of incorporating a bias weighting factor is that it can be used to guide search towards the discovery of hypotheses in which the emphasis may be on achieving either a low false negative rate, or a low false positive rate. This is a useful property, because in many application domains misclassification costs are not equal. Although determining an appropriate bias weighting may require some degree of subjective judgement, heuristics are presented to assist in this task. The technique is able to cope with noise in the training set, and is applicable to a broad range of classification and pattern recognition problems.