Fuzzy Inference Methods Applied to the Learning Competence Measure in Dynamic Classifier Selection

The concept of classifier competence in the feature space is fundamental to dynamic classifier selection in multiple classifier systems (MCS). Competence function (measure) of base classifier can be determined using validation set in the two step procedure. The first step consists in creating competence set, i.e. the set of classifier competences for all validation objects. To this end a hypothetical classifier called randomized reference classifier (RRC) is constructed. Since RRC - on average - acts like the evaluated classifier, the competence of the classifier at validation point is calculated as the probability of correct classification at this point of the respective RRC. In the second step, the competences calculated for a validation set are generalised to an entire feature space by constructing a competence function based on a supervised learning procedure. In this study, the second step of the above procedure is addressed by developing the fuzzy inference methods of learning competence functions. Two fuzzy inference systems are developed and applied to the supervised learning competence function of base classifiers in a MCS system with dynamic classifier selection (DCS) and dynamic ensemble selection (DES) scheme: Mamdani fuzzy inference system and Sugeno fuzzy inference system. Both fuzzy inference systems were experimentally tested and compared against 4 literature methods of learning classifier competence (potential function, regression model, multilayer perceptron, k-nearest neighbor scheme) using 9 databases taken from the UCI Machine Learning Repository. The experimental results clearly show the effectiveness of the proposed supervised learning competence function using fuzzy inference systems regardless of the ensemble type used (homogeneous or heterogeneous).