Predicting gambling addiction using knowledge graph and machine learning

Abstract

The aim of this study is to apply machine learning and knowledge graph technologies to the online gambling domain, to help predicting players at risk of developing gambling addiction. While previous research would often focus exclusively on Voluntary Self-Exclusion, we believe that the usage of the Voluntary Self-Exclusion tool is not always a good proxy to predict gambling addiction, as players might use this tool as a quick way to close their accounts. In our study we used features related to the customers’ behaviour based on the Diagnostic and Statistical Manual of Mental Disorders fifth edition (DSM-5), to find the most accurate technique between five machine learning models (random forest, naive Bayes, gradient boost, logistic regression, K-NN) and five Knowledge Graph Embedding models (TransE, TransR, DistMult, ComplEx, RotatE). We also investigated the use of techniques to balance our dataset, as our at-risk players were only around 1% of the population. Furthermore, we experimented with the Shapley Additive Explanations technique, to understand the reasons behind the detection. We also investigated how well gambling addiction predictions works in Graphs. More specifically, we built our customers’ behaviour Knowledge Graph, called BKG, and implemented Knowledge Graph Embedding models, so that we could find similarities between players who used Voluntary Self-Exclusion tools and players who did not use Voluntary Self-Exclusion tools at the time of this study. Finally, we used the embeddings of ComplEx as input in the gradient boosting model. We concluded that using DSM-5 features for machine learning and Knowledge Graph Embedding models is a promising approach. Our best machine learning model was gradient boosting, which achieved an AUC-ROC of 0.86. For our Knowledge Graph Embedding experiments our best model was ComplEx with MR of 1.86. The experiment to check players with similar features also performed well, as we were able to detect around 66% of the players who had not used the Voluntary Self-Exclusion at the time of the study, but would end up using it after the data had been collected. On the other hand, our approach of using embeddings as input in the gradient boosting model did not perform as expected, as the most accurate model was still gradient boost with the tabular data.