Meent & Tomaszewski - Fusing Machine Learning and Mixed Integer Linear Programming

Learn how to combine machine learning with mixed integer linear programming to optimize marketing spend, while accounting for real-world constraints and uncertainty.

Key takeaways
  • Machine learning models combined with mixed integer linear programming help optimize marketing spend allocations and promotional campaigns

  • Models need regular monitoring and retraining to account for shifts in market conditions, competitor behavior, and customer responsiveness

  • Key challenges include:

    • Limited historical data makes effects hard to measure
    • Correlations between variables can cause instability
    • Model uncertainty needs to be considered in optimization
    • Changes in external factors (weather, competition, etc.) affect predictions
  • Best practices for implementation:

    • Validate models before deploying to optimization
    • Use randomized experiments to measure true causal effects
    • Consider business constraints beyond mathematical optimality
    • Monitor key performance metrics and model drift
    • Maintain awareness of model limitations and uncertainty
  • Response curves help identify “sweet spots” for marketing investments and identify when diminishing returns set in

  • Regularization and monotonicity constraints can help produce more stable and interpretable models

  • Combining human knowledge with model outputs leads to better decision-making than pure optimization

  • Important to design robust solutions that work well across different scenarios rather than solely mathematically optimal ones

  • Regular model updates and monitoring help adapt to changing market conditions and maintain effectiveness

  • Critical to validate that model recommendations make business sense before implementation