ModBoost for unbiased classification

Many real world data mining applications involve learning from large data sets. Boosting, an ensemble-based learning algorithm, has shown to improve the performance of classifiers in many situations. In this paper, we describe a new approach that is a modified version of boosting. In the ModBoost method, the sample training set is produced from the given dataset by sampling with replacement. The model is generated from the sample training set and is evaluated for error on the dataset instead of the training set. Misclassified tuples are given a higher weight and subsequently, in the next iterations the misclassified tuples are given first preferences and resampling is done for the rest of the training set. The total weights of the different classes in the new training set are rebalanced. The ModBoost method is evaluated, in terms of the F-measures, Precision, Recall and overall accuracy against large and moderately imbalanced bank marketing data set using Naïve Bayesian as a base classifier. The results are promising and show that the ModBoost method compares well in comparison with AdaBoost and bagging algorithm. Our results prove empirically that ModBoost is an attractive alternative for improving the classification performance. We also provide suggestions for future research.