CDBH: A clustering and density-based hybrid approach for imbalanced data classification

摘要

The problem of imbalanced data set classification is prevalent in the studies of machine learning and data mining. In these kinds of data sets, the number of samples in classes is unequal so that one class has a lot more samples (the majority or negative class) than the other (the minority or positive class). The classical classifiers are ineffective in these conditions because they are biased toward the majority class and ignore the minority class, which is more important. Preprocessing the data distribution before training the classifier is one of the most effective methods to resolve this problem. These methods, balance the data distribution by decreasing the majority class size (under-sampling methods) or increasing the minority class size (over-sampling methods) or combining both of them (hybrid methods). In this paper, we propose an effective and simple hybrid approach based on the density concept and clustering, which is called Clustering and Density-Based Hybrid (CDBH). First, the minority class samples are clustered by the well-known k-means algorithm and their densities in each cluster are obtained. Then, the denser minority samples are selected with more likely to generate the new minority samples. To decrease the majority class size, the k-means algorithm is applied again on the majority class samples to cluster them and compute their densities, like the previous stage. Finally, the denser majority samples will have more chance to choose from the training set, and other samples are removed to balance the data samples distribution between classes. In the experiments, the Support Vector Machine (SVM) classifier is used as the classifier, and F-measure and AUC criteria are employed for evaluation. Also, preprocessing methods are compared in terms of the complexity of the classification model and the over-sampling rate. The results of comparing CDBH and other state of the art methods over 44 imbalanced data sets show the superiority of the proposed CDBH method based on the F-measure criterion.