Accuracy and interpretability trade-offs in machine learning applied to safer gambling

Abstract

Responsible gambling is an area of research and industry which seeks to understand the pathways to harm from gambling and implement programmes to reduce or prevent harm that gambling might cause. There is a growing body of research that has used gambling behavioural data to model and predict harmful gambling, and the industry is showing increasing interest in technologies that can help gambling operators to better predict harm and prevent it through appropriate interventions. However, industry surveys and feedback clearly indicate that in order to enable wider adoption of such data-driven methods, industry and policy makers require a greater understanding of how machine learning methods make these predictions.
In this paper, we make use of the TREPAN algorithm for extracting decision trees from Neural Networks and Random Forests. We present the first comparative evaluation of predictive performance and tree properties for extracted trees, which is also the first comparative evaluation of knowledge extraction for safer gam- bling. Results indicate that TREPAN extracts better performing trees than direct learning of decision trees from the data. Overall, trees extracted with TREPAN from different models offer a good compromise between prediction accuracy and interpretability. TREPAN can produce decision trees with extended tests rules of different forms, so that interpretability depends on multiple factors. We present detailed results and a discussion of the trade-offs with regard to performance and interpretability and use in the gambling industry.

Problem with this document? Please report it to us.