Customer Value Prediction in Direct Marketing Using Hybrid Support Vector Machine Rule Extraction Method

Abstract

Data mining techniques can aid companies in evaluation of customers that generate highest amount of revenue in a direct marketing campaign. Most commonly, customer value is evaluated by a uniform segmentation of customers (20% for each segment) based on buying behavior using recency, frequency and monetary (RFM) attributes, whereby for direct campaigns the segments with the highest score of these attributes are subjectively selected. In this paper, the method of k-means clustering, according to RFM attributes is proposed, based on which the customer value can be more objectively determined. The most valuable customers, as a rule, are the smallest group compared to other clusters, so the problem of class imbalance occurs. In order to overcome this problem, a hybrid Support Vector Machine Rule Extraction (SVM-RE) method is proposed for predicting which customer belongs to a cluster, based on data on consumer characteristics and offered products. The SVM classifier is known as a good predictor in case of class imbalance, but does not generate an interpretable model. Therefore, the Decision Tree (DT) method generates rules, based on the prediction result of the SVM classifier. The results of the empirical case study showed, that using this hybrid method with good classification performance, customer value level can be predicted, i.e. targeting existing and new buyers for direct marketing campaigns can be efficiently done, regardless of the class imbalance problem. It’s also shown that using the hybrid SVM-RE method, it is possible to obtain significantly better prediction accuracy than using the DT method.