Evidential Reasoning combine with Mass-based Similarity for Imbalanced Classification

This paper proposes an integrated approach based on Demspter-Shafer theory of evidence and mass-based dissimilarity measurement for imbalanced classification problem. We aim to overcome the weaknesses of distance-based and density-based classifiers when there is an unequal distribution of classes in datasets. There are two main challenges: (1) the misclassification bias due to the equal treatment for all instances when using conventional classifiers, although the majority classes contain most of these instances in the skewed datasets; (2) the varied densities of instances in datasets. The proposed classifier so-called EMass utilizes new mass- based measures instead of using distance-based or density-based calculations to assess the similarity between objects. In addition, we consider each neighbor of the query instance as a prior information source that provides evidence to support the classification task. Each piece of evidence is formulated through the so-called basic probability assignment, then the Dempster’s rule of combination is applied to combine multi-pieces of evidence to obtain the overall probability for making decision. We conduct experiments and use evaluation metrics such as the F1 score, the Brier score, and the area under the ROC curve values to compare 12 tested models performing on 60 imbalanced datasets. Wilcoxon signed ranks analysis is also conducted to validate all the pairwise correlation results.