Under the Spotlight: Understanding the Risk Profile of Gambling Behaviour through Machine Learning Predictive Modelling and Explanation

The document likely discusses using machine learning models to predict and explain risky gambling behaviour.

Categories

This research introduces a novel method—feature risk curves—to explain how individual behavioural features influence predictions made by machine learning models used in gambling harm prevention. The method was applied to a real-world system developed by Playtech’s BetBuddy to identify players at risk of gambling harm.

Key Findings:

Practical Use in Safer Gambling: The approach supports model validation, regulatory transparency, and the development of more nuanced player interventions. It also highlights the limitations of using self-exclusion as a sole proxy for harm.

Feature Risk Curves Offer Directional Insight: Unlike traditional feature importance rankings, feature risk curves visualize how changes in a single feature (e.g. night-play ratio) affect the model’s predicted risk, while holding other variables constant. This helps experts understand not just which features matter, but how they matter.

Night-Play Ratio Shows Non-Linear Risk: Players with moderate night-time play (38–68%) had the highest predicted risk of serious self-exclusion. However, those with very high night-play ratios (94%+) showed lower risk, possibly reflecting stable routines (e.g. shift workers) rather than problematic escalation.

Deposit Decline Ratio Not Strongly Predictive: Contrary to industry assumptions, frequent declined deposits did not significantly increase predicted risk. This suggests that deposit declines alone may not be a reliable indicator of harm without context from other behaviours.

Model Fidelity and Explainability: The method is model-agnostic and scalable, relying on querying the original ML model directly (e.g. a random forest trained on 41 behavioural features). It avoids the pitfalls of surrogate models and provides high-fidelity, interpretable outputs.

Practical Use in Safer Gambling: The approach supports model validation, regulatory transparency, and the development of more nuanced player interventions. It also highlights the limitations of using self-exclusion as a sole proxy for harm.

Conclusion:

Feature risk curves provide a powerful tool for interpreting complex ML models in gambling harm prevention. They enable domain experts to explore the directional impact of behavioural features and challenge simplistic assumptions, paving the way for more targeted and effective responsible gambling strategies.

For full details, please refer to the complete document.