New Horizons in Responsible Gambling: Making BlackBox Algorithms Understandable

BetBuddy Lead Researcher explains efforts to make complex machine learning algorithms predicting gambling harm understandable.

Categories

The presentation investigates the interpretability of machine learning algorithms used to predict self-exclusion in online gambling, highlighting the trade-offs between accuracy and transparency. Key findings:

• Four ML Techniques Compared: Logistic regression, Bayesian networks, neural networks, and random forest were applied to behavioral data from 845 gamblers. Random forest achieved the highest accuracy (87%) but was the least interpretable.
• Five Risk Factors Identified: Frequency, trajectory, intensity, variability, and session time were used to model gambling behavior. These were derived through human-led pre-processing, emphasizing the importance of domain expertise.
• Interpretability vs Accuracy Trade-Off: Complex models like neural networks and random forests offer better predictive performance but are difficult to explain. Simpler models (e.g., logistic regression) are easier to interpret but less accurate.
• TREPAN for Model Simplification: TREPAN was used to convert neural networks into decision trees, retaining 87% fidelity with minimal accuracy loss, offering a pathway to more interpretable outputs.
• Individual-Level Explanation Still Lacking: While model-level insights are possible, explaining why a specific gambler was flagged remains challenging. The presentation calls for layered interpretative tools to bridge this gap.

These insights support the development of responsible gambling tools that balance predictive power with transparency, enabling both industry-level action and individual-level understanding. For full details, please refer to the complete presentation.