SSPO's structure-aware weighting boosts neural optimization, now live at JD.com

Neural combinatorial optimization (NCO) trains models to solve routing, scheduling, and facility-location problems by sampling many candidate solutions in parallel. But existing training methods waste information hidden in the co-sampled solution group. Preference-optimization approaches anchor only on the single best solution, discarding fine-grained quality and structural signals from all other peers—a failure the authors call gradient signal polarization. Mean-based baselines, meanwhile, weight all peers uniformly, so structurally similar solutions flood the baseline with redundant data and keep gradient variance high—a failure called baseline redundancy.

SSPO fixes both problems with one mechanism: a dissimilarity-weighted leave-one-out baseline that scores all B sampled solutions jointly, giving structurally distinct peers higher weight. The baseline relies on zero-parameter, problem-adaptive solution embeddings built from the encoder's existing node representations, so it adds no learnable overhead. Experiments on TSP (traveling salesman), EFL (facility location), and JSP (job shop scheduling) benchmarks show consistent gains over both best-anchor and uniform-weight baselines. Ablations confirm that structure-aware weighting is the primary driver of improvement, not the leave-one-out trick. The SSPO-trained EFL policy has already been deployed in a production facility-location system at JD.com, demonstrating that the method works beyond academic benchmarks.