Predicting individual behaviour: an empirical approach in online marketing

Abstract

This paper investigates the use and relevance of data mining techniques for online direct marketing. Cookie log files are obtained and transformed into time-aggregated web user characteristics to predict which users are likely to purchase. Given these characteristics, users observe relevant banners. Modern classification techniques, i.e., support vector machines (SVM), random forests (RF), bagging (BA), and boosting (BO), are compared with classic data mining techniques, i.e., multinomial logistic (MNL) regressions, neural networks (NN), and Naive Bayes (NB). We found that, after feature selection, all modern techniques significantly outperform the classic methods NN and NB, accuracy-wise. MNL performs similarly to BO and better than BA. RF performs best with an average accuracy of 70.7%, followed by SVM with an average accuracy of 67.6%. The RF model has led to a decrease in banners served, while preserving the number of sales. Several novel time-related features have also been proposed for online bannering.