JD's DEGR re-ranking model lifts e-commerce CTR by 1.22%

In industrial recommendation systems, the re-ranking stage typically optimizes sequence-level business objectives within a fixed candidate pool. However, when upstream supply is low-quality, these methods hit a ceiling—they cannot improve effectiveness because they only reorder what's already been retrieved. DEGR breaks this constraint by actively balancing immediate reward (e.g., click-through) with exploratory value, intentionally surfacing less certain items to preserve user browsing potential and unlock serendipitous conversions. The paper, accepted at KDD 2026 ADS Track, describes DEGR as a generator that learns a kind of adaptive cross-request contextual bridge, carrying exploration decisions across requests.

The method's core is a hybrid optimization paradigm combining supervised learning, an exploration diversity constraint, and adaptive reward-weighted ORPO for preference optimization. This dual exploration—both at the reward-model level and in the generator's training—lets DEGR shift between exploitation (when supply is rich) and exploration (when it's scarce) without explicit rules. Offline and online experiments on JD's platform show DEGR consistently beats state-of-the-art baselines, with up to 1.22% improvement in UCTR (user click-through rate) and 0.20% in PV (page views). These are meaningful gains for a real large-scale e-commerce system, suggesting that re-ranking can evolve from a passive reorderer into an active policy that manages user engagement across requests.