Search papers, labs, and topics across Lattice.
This paper introduces the DEGR method, which employs a hybrid supervised-reinforcement learning approach to optimize the re-ranking stage in recommendation systems by balancing immediate and exploratory value. By prioritizing exploratory exposure, DEGR effectively enhances browsing potential and facilitates serendipitous conversions, particularly in scenarios with low-quality supply. Experimental results demonstrate that DEGR significantly outperforms state-of-the-art methods, achieving notable improvements in user click-through rates (UCTR) and page views (PV) within the JD E-commerce recommendation system.
Balancing immediate and exploratory value in recommendation systems can lead to a 1.22% boost in user engagement, even under low-quality supply conditions.
In industrial recommendation systems, the re-ranking stage balances business objectives and diversity for sequence-level optimization while modeling contextual information. However, constrained by fixed upstream supply, existing methods fail to deliver further effectiveness gains, especially under low-quality supply. To overcome this, re-ranking can actively balance immediate and exploratory value, for instance, by prioritizing exploratory exposure under low-quality supply to preserve browsing potential and facilitate serendipitous conversions. Therefore, we propose a Dual Exploration-Driven Generative Re-Ranking (DEGR) method. DEGR adopts a hybrid supervised-reinforcement exploration and optimization paradigm, guided by an exploratory reward model that adaptively balances immediate and exploratory value. The hybrid optimization paradigm integrates three key components: supervised learning, exploration diversity constraint, and adaptive reward-weighted ORPO for preference optimization. Through this dual exploration, the generator ultimately acts as an adaptive cross-request contextual bridge. Offline and online experiments indicate that DEGR outperforms SOTA methods, achieving improvements of up to 1.22% UCTR and 0.20% PV in the JD E-commerce recommendation system.
