Handling Imbalanced Data through Affinity Propagation and SMOTE
Lanie B. Laureano, Ariel M. Sison, Ruji P. Medina · 2019
Imbalanced datasets occur in real-world scenarios and affect the performance of the classifiers in supervised learning. To combat the imbalance between majority and minority instances, several techniques including SMOTE which generates artificial data along the line between minority samples and its selected neighbors are used in preprocessing. However, SMOTE suffers from generating synthetic data in noise regions and imbalanced distribution of samples within the minority class. Using clustering techniques prior to the oversampling step to determine the target areas of the input space where the generation of synthetic samples is effective, is known to be an extension of SMOTE. The proposed method uses affinity propagation as a clustering technique to generate clusters and its cluster exemplar without requiring the specification of the number of clusters a priori. The cluster exemplar becomes the basis of which samples are to be oversampled in each cluster. Simulation results on some publicly available datasets show the effectiveness of the proposed method in terms of F-measure, G-mean and AUC with a mean value across datasets of .685, .756 and .765 respectively.