Contextual Decision Making: Predictive or Generative

TL;DR

This paper compares predict-then-optimize (PtO) and generate-then-optimize (GtO) for contextual decision making, showing that PtO's simplicity can outperform GtO when smoothness in the problem structure is leveraged, despite GtO's richer modeling of uncertainty. The study establishes regret bounds that reveal tradeoffs between statistical accuracy, optimization tractability, and convexity preservation. Real and synthetic experiments demonstrate context-dependent superiority based on data size and problem structure.

Why It Matters

The findings offer practical guidance for choosing between two dominant paradigms in decision-making under uncertainty, helping practitioners avoid overcomplicating models when simpler approaches suffice. By clarifying when smoothness or scenario generation provides real benefits, the work supports more efficient and reliable AI-driven decision systems in operations research and machine learning.

Key Findings

• PtO can outperform GtO when the problem exhibits sufficient smoothness, despite learning a narrower target.
• Smoothness enables better regret bounds in PtO, but there is a fundamental tradeoff: increased smoothness offers diminishing returns and may require signed weights that break convexity.
• GtO preserves scenario-wise convexity but demands learning a richer conditional distribution and generating scenarios, making it more data-hungry and computationally intensive.