A new oversampling method and improved radial basis function classifier for customer consumption behavior prediction

In practical applications, imbalanced data has brought great challenges to classification problems. In this paper, we propose two new methods: (1) a new oversampling method, named Tomek-CASUWO, to address the issue of class imbalance; (2) a new classifier, named ILS-RBFNN, to increase the accuracy of prediction of customer consumption behavior. Since current ASUWO algorithm is easily affected by noise samples and fuzzy class boundaries, we propose Tomek-CASUWO to address these problems: (i) the Tomek links algorithm is used to filter noise samples; (ii) CASUWO is used to avoid overlapping class boundaries; (iii) Tomek-CASUWO is used to synthesize new samples. Also, we propose a new classifier based on RBFNN, named ILS-RBFNN, to improve the prediction accuracy: (i) the hybrid kernel is developed by combining Gaussian and Polynomial; (ii) an Immune Algorithm (IA) is used to optimize the centers of RBFNN; (iii) Least-Squares (LS) is used to calculate the biases and weights. Wine-consumer behavior data is used to compare our Tomek-CASUWO with other oversampling methods. We compare ILS-RBFNN with several well-known kernel functions and parameter update methods. The experimental results show that Tomek-CASUWO can significantly improve the prediction accuracy of a classifier, and ILS-RBFNN outperforms other classification methods. We also conduct experiments on the extended real-world dataset. Finally, the robustness and applicability of ILS-RBFNN are verified on the eleven UCI datasets. All results show that the proposed two methods outperform existing models. The experimental results also demonstrate the effectiveness and practicability of ILS-RBFNN for predicting customer behavior.