Application of multilayer perceptrons for response modeling

This study explores the predictive abilities of multilayer perceptrons used for response modeling in direct marketing campaigns. We extend previous studies discussing how neural network design affects the model performance and also propose a simplified architecture, which outperforms the ones used before. We explore the variance in the neural network behaviour due to the randomness factor and validate the figures of merit by a rigorous testing procedure not applied in the previous studies. The model performance is estimated and analyzed using accuracy, ROC, AUC, lift, and precision-recall. We also compare multilayer perceptrons with logistic regression, naive Bayes, linear discriminant analysis, and quadratic discriminant analysis.