Balancing Competing Objectives in Generative Recommendation

Generative recommender systems often struggle to satisfy multiple, conflicting objectives simultaneously—such as maximizing user engagement while ensuring diversity, fairness, or business-specific constraints. Traditional approaches often rely on weighted sum objectives during training, which are rigid and fail to adapt to dynamic constraint requirements at inference time. The authors propose a Stochastic Primal-Dual Decoding (SPDD) framework that treats recommendation as a constrained optimization problem solved during the decoding process.

The Primal-Dual Decoding Mechanism

Instead of baking constraints into the model weights, SPDD introduces a dual variable update mechanism that adjusts the decoding probability distribution dynamically.

• Primal Step: The model generates candidate items based on the current policy, influenced by the dual variables (Lagrange multipliers) that represent the 'cost' of violating specific constraints.
• Dual Step: The system updates these multipliers based on the observed constraint violations in the generated output. If a constraint (e.g., minimum diversity threshold) is violated, the dual variable increases, effectively penalizing the model for selecting items that contribute to that violation in the next step.

This approach allows the system to enforce hard constraints on metrics like novelty, category coverage, or fairness without requiring expensive model fine-tuning. By performing these updates stochastically during inference, the system remains computationally efficient while providing a principled way to navigate the trade-off space between relevance and secondary objectives.