Adaptive Multi-objective Swarm Crossover Optimization for Imbalanced Data Classification

摘要

Training a classifier with imbalanced dataset where there are more data from the majority class than the minority class is a known problem in data mining research community. The resultant classifier would become under-fitted in recognizing test instances of minority class and over-fitted with overwhelming mediocre samples from the majority class. Many existing techniques have been tried, ranging from artificially boosting the amount of the minority class training samples such as SMOTE, downsizing the volume of the majority class samples, to modifying the classification induction algorithm in favour of the minority class. However, finding the optimal ratio between the samples from the two majority/minority class for building a classifier that has the best accuracy is tricky, due to the non-linear relationships between the attributes and the class labels. Merely rebalancing the sample sizes of the two classes to exact portions will often not produce the best result. Brute-force attempt to search for the perfect combination of majority/minority class samples for the best classification result is NP-hard. In this paper, a unified preprocessing approach is proposed, using stochastic swarm heuristics to cooperatively optimize the mixtures from the two classes by progressively rebuilding the training dataset is proposed. Our novel approach is shown to outperform the existing popular methods.