The effect of varying levels of class distribution on bagging for different algorithms: An empirical study

Many real world applications involve highly imbalanced class distribution. Research into learning from imbalanced class distribution is considered to be one of ten challenging problems in data mining research, and it has increasingly captured the attention of both academia and industry. In this work, we study the effects of different levels of imbalanced class distribution on bagging predictors by using under-sampling techniques. Despite the popularity of bagging in many real-world applications, some questions have not been clearly answered in the existing research, such as the effect of varying the levels of class distribution on different bagging predictors, e.g., whether bagging is superior to single learners when the levels of class distribution change. Most classification learning algorithms are designed to maximize the overall accuracy rate and assume that training instances are uniformly distributed; however, the overall accuracy does not represent correct prediction on the minority class, which is the class of interest to users. The overall accuracy metric is therefore ineffective for evaluating the performance of classifiers in extremely imbalanced data. This study investigates the effect of varying levels of class distribution on different bagging predictors based on the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) as a performance metric, using an under-sampling technique on 14 data-sets with imbalanced class distributions. Our experimental results indicate that Decision Table (DTable) and RepTree are the learning algorithms with the best bagging AUC performance. The AUC performances of bagging predictors are statistically superior to single learners, with the exception of Support Vector Machines (SVM) and Decision Stump (DStump).