Laura Israel - Je ne regrette rien: Teaching Machine Learning Models Regret Avoidance

Learn how to optimize ML systems by incorporating regret avoidance into model training through custom loss functions and sample weights, improving sequential decision-making outcomes.

Key takeaways
  • Complex ML systems often involve multiple sequential decision components (models, rules, human decisions) that contribute to the same higher-level goal

  • Regret in ML systems occurs when decisions turn out to be suboptimal after more information becomes available - this can be incorporated into model training

  • Sample weights and custom loss functions can be used to penalize decisions that lead to regret conditions, without replacing existing models

  • The regret parameter should be tuned carefully - if set too high it can lead to underperformance both locally and globally

  • Implementation demonstrated using XGBoost by modifying the loss function and applying sample weights to regret cases

  • Key benefits include:

    • Simple implementation with existing ML frameworks
    • No need for complex model changes
    • Can improve global optimization while maintaining local performance
  • Results showed significant improvements - in one case converting 300,000 more transactions per month worth €20M in additional merchant revenue

  • Approach works best when:

    • System has sequential decision points
    • Later information reveals optimal decisions
    • Components contribute to shared objectives
    • Regret conditions can be clearly defined
  • Important considerations:

    • Regret parameter needs careful tuning
    • Should be stable across optimization
    • Implementation depends on use case
    • May need to adapt for continuous targets
  • Complex systems can be viewed as game trees with imperfect information games, allowing application of game theory concepts