Improved Imbalanced Classification with CCCD and its Variants

Abstract

We use a graph based classification method called class cover catch digraphs (CCCDs) to solve class cover problem. CCCD is a random graph model gives graph solution to calss cover problem and shows relatively good performance in class imbalance problem. We will focus on the improvement of CCCD performance. Ensemble methods bagging and boosting will be employed to modify CCCD. We show and compare the performance of the ensemble CCCD classifiers with other commonly used classifiers by Monte Carlo simulation analysis. From the results, CCCDs are very robust to imbalanced data. We tested ensemble with Different Parameters. When dealing with local imbalanced data, ensembling can slightly improve the the performance of P-CCCD and RW-CCCD in all dimensions setting while E-Comb showed the best performance among all classifiers. When dealing with local balanced data, ERW-CCCD showed slightly better performance in low dimensions. Both EP-CCCD and ERW-CCCD showed poor performance in high dimensions but E-Comb showed better performance when dimension increased. E-Comb also showed a relatively stable performance among all methods. Bagging can significantly improve the performance of CCCD when dealing with local imbalanced dataset but it does not show much improvement when dealing with local balanced dataset.