A Measure Optimized Cost-Sensitive Learning Framework for Imbalanced Data Classification

Abstract

Class imbalance is one of the challenging problems for machine learning in many real-world applications. Many methods have been proposed to address and attempt to solve the problem, including sampling and cost-sensitive learning. The latter has attracted significant attention in recent years to solve the problem, but it is difficult to determine the precise misclassification costs in practice. There are also other factors that influence the performance of the classification including the input feature subset and the intrinsic parameters of the classifier. This paper presents an effective wrapper framework incorporating the evaluation measure (AUC and G-mean) into the objective function of cost sensitive learning directly for improve the performance of classification, by simultaneously optimizing the best pair of feature subset, intrinsic parameters and misclassification cost parameter. The optimization is based on Particle Swarm Optimization (PSO).We use two different common methods, support vector machine and feed forward neural networks to evaluate our proposed framework. Experimental results on various standard benchmark datasets with different ratios of imbalance and a real-world problem show that the proposed method is effective in comparison with commonly used sampling techniques. INTRODUCTION Recently, the class imbalance problem has been recognized as a crucial problem in machine learning and data mining (Chawla, Japkowicz K Kotsiantis, Kanellopoulos& Pintelas, 2006; He G He & Ma, 2013). This issue of imbalanced data occurs when the training data is not evenly distributed among classes. This problem is also especially critical in many real applications, such as credit card fraud detection when fraudulent cases are rare or medical diagnoses where normal cases are the majority, and it is growing in importance and has been identified as one of the 10 main challenges of Data Mining (Yang, 2006). In these cases, standard classifiers generally perform poorly. Classifiers usually tend to be overwhelmed by the majority class and ignore the minority class examples. Most classifiers assume an even distribution of examples among classes and assume an equal misclassification cost. Moreover, classifiers are typically designed to maximize accuracy, which is not a good metric to evaluate effectiveness in the case of imbalanced training data. Therefore, we need to improve traditional algorithms so as to handle imbalanced data and choose other metrics to measure performance instead of accuracy. We focus our study on imbalanced datasets with binary classes. Much work has been done in addressing the class imbalance problem. These methods can be grouped in two categories: the data perspective and the algorithm perspective (He &Garcia 2009). The methods with the data perspective re-balance the class distribution by re-sampling the data space either