Application of Data Mining Techniques for Direct Marketing

This paper presents a case study of data mining modeling techniques for direct marketing. It focuses to three stages of the CRISP-DM process for data mining projects: data preparation, modeling, and evaluation. We address some gaps in previous studies, namely: selection of model hyper-parameters and controlling the problem of under-fitting and over-fitting; dealing with randomness and 'lucky' set composition; the role of variable selection and data saturation. In order to avoid overestimation of the model performance, we applied a double-testing procedure, which combines cross-validation, multiple runs over random selection of the folds and hyper-parameters, and multiple runs over random selection of partitions. The paper compares modeling techniques, such as neural nets, logistic regression, naive Bayes, linear and quadratic discriminant analysis, all tested at different levels of data saturation. To illustrate the issues discussed, we built predictive models, which outperform those proposed by other studies.