SSPO: Structure-Aware Similarity-Weighted Preference Optimization for Neural Combinatorial Optimization

2026-08-12

AI Summary

Existing neural combinatorial optimization methods struggle to effectively leverage structural and quality information from co-sampled solutions during parallel sampling training, often leading to gradient polarization or redundant baselines. This work proposes a structure-aware similarity-weighted preference optimization mechanism that constructs zero-parameter, problem-adaptive solution embeddings from encoder node representations, estimates baselines via leave-one-out evaluation, and assigns higher weights to solutions with greater structural dissimilarity. By integrating structure-aware weighting into the preference optimization framework for the first time, the approach jointly mitigates both gradient polarization and baseline redundancy. The method consistently outperforms state-of-the-art approaches on TSP, EFL, and JSP benchmarks, with the EFL policy already deployed in JD.com’s production facility location system, demonstrating its practical scalability and real-world applicability.

Abstract

Neural combinatorial optimization (NCO) relies on parallel solution sampling for training, yet existing methods fail to fully exploit the rich information latent in a co-sampled solution group. Preference-optimization methods anchor on the single best solution and discard fine-grained quality and structural signal from all other peers-a failure we term gradient signal polarization. Mean-based baselines instead weight peers uniformly, so structurally near-identical peers flood the baseline with redundant information and keep gradient variance high-a failure we term baseline redundancy. We propose SSPO (Structure-Aware Similarity-Weighted Preference Optimization), which scores all $B$ sampled solutions jointly through a dissimilarity-weighted leave-one-out baseline: structurally distinct peers receive higher weight, resolving both failures in a single mechanism. The baseline uses zero-parameter, problem-adaptive solution embeddings built from the encoder's existing node representations. Experiments on TSP, EFL, and JSP benchmarks show consistent gains over prior best-anchor and uniform-weight baselines. A direct comparison against uniform RLOO on TSP and EFL confirms that structure-aware weighting is the primary driver of improvement. The SSPO-trained EFL policy has been deployed in a production facility-location system at JD$\mathord{.}$com, confirming practical viability at scale.

Problem

• Neural Combinatorial Optimization
• Preference Optimization
• Gradient Signal Polarization
• Baseline Redundancy
• Solution Sampling

Innovation

• Structure-Aware
• Similarity-Weighted
• Preference Optimization
• Neural Combinatorial Optimization
• Leave-One-Out Baseline