Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches (NIPS 2016)

BetBuddy and City AI research explored machine learning model interpretation techniques for safer gambling at NIPS 2016.

Categories

This paper explores how machine learning can be used to predict harmful gambling behavior while balancing accuracy and interpretability. It introduces the TREPAN algorithm as a method for extracting decision trees from complex models like neural networks and random forests, making them more understandable to industry stakeholders.

Key Insights:

• Industry Need:
  Gambling operators and regulators prioritize interpretability over pure accuracy. Surveys show that 70% of stakeholders prefer a model that is less accurate but more transparent.
• Data & Models:
  The study uses behavioral data from IGT in Ontario, Canada, including 13,615 control players and 449 self-excluders. Models were built using random forests and neural networks, with additional features engineered to capture loss-related behaviors.
• Performance Comparison:
  • Random Forest: 90% accuracy
  • Neural Network: 84% accuracy
  • Standard Decision Tree: 76% accuracy
    TREPAN-generated trees showed slightly reduced accuracy but improved interpretability.
• TREPAN Algorithm:
  TREPAN extracts decision trees using various rule structures (M-of-N, N-of-N, 1-of-N, 1-of-1).
  • M-of-N trees offer high accuracy but are harder to interpret.
  • 1-of-N trees from neural networks provided the best balance of accuracy (up to 85%) and interpretability.
• Interpretability vs. Accuracy Trade-off:
  TREPAN trees reduce model complexity while maintaining acceptable accuracy (0–7% loss). Smaller trees are easier to interpret and fit on a single page, unlike random forests which can span thousands of pages.
• Implications for Responsible Gambling:
  Transparent models can support better player protection, regulatory compliance, and therapeutic interventions. They also enable personalized messaging and early warnings for at-risk players.

TREPAN offers a practical solution for extracting interpretable models from complex machine learning systems in gambling. Future research should focus on optimizing tree structures and evaluating interpretability with domain experts.

For full insights please refer to the complete document.