Machine learning techniques for the determination and prediction of online gambling addiction

Abstract

This project aims to provide a reliable software application that helps to predict potential problematic online gambling behaviour through the use of multiple machine learning techniques. This project answers the following research questions with regards to online gambling addiction: (1) Which are the techniques or types of methods that offer the most significant potential for solving this problem? (2) How stable and accurate are these methods in predicting problematic gambling? The literature review had an essential part in this project to examine as many previous studies published on this particular problem as possible including those related to similar tasks from other areas of studies, to have a broad number of viable techniques to evaluate. This experimental study adopted a mixed methodology, in which both qualitative and quantitative methods were used to clean and analyse the dataset obtained through The Transparency Project (a project collaboration between the Division on Addiction and bwin.party). In turn, this was the data used to train, validate, and evaluate all selected models, combined with stratified 10-fold cross-validation to use all data available for all phases of implementation. In conclusion, boosting methods like AdaBoost, LightGBM, and XGBoost and other ensemble techniques such as Random Forest, Extremely Randomised Trees and Regularised Greedy Forest were found to be the most useful implementations (out of the evaluated) in predicting possible cases of online gambling addicting. All of these methods achieved a mean score higher than 80% across all four performance metrics assessed: Area Under the Receiver Operating Characteristic, Accuracy, Sensitivity, and Specificity. Moreover, due to time being a limiting factor, further evaluation of additional techniques, as well as improvements in the application itself, should be part of future works.