Rule-based classification by means of bipolar criteria

Classification problems often play an important role in many decision contexts. Therefore, the design of decision support tools to operate in such contexts usually involves the formulation of adequate classification models. Fuzzy rule-based classifiers FRBCS are excellent methodological tools for this purpose due to their interpretability and ability to deal with linguistic knowledge representations. Learning of these rules from data is an increasingly common practice in order to avoid complex knowledge engineering processes. This paper proposes the notions of minor and significant exceptions to a rule in order to extend the notion of counterexample and thus enhance the representational and modelling power of FRBCS. This allows to consider some classes as being dissimilar or opposite, and leads to the introduction of a bipolar approach in rule based learning for classification, as the evaluation of rules in terms of positive and negative evidence is enabled in this way. As a consequence, it is then possible to introduce significant features and requirements of the decision contexts in the underlying classification models in a flexible and practical way. In order to illustrate the usage of the proposed bipolar classification framework, an example of application in the context of humanitarian logistics decision making is described.