UniRec: Bridging the Expressive Gap between Generative and Discriminative Recommendation via Chain-of-Attribute
New generative recommendation model outperforms baselines by 22.6% by adding structured item attributes before decoding.
A research team has introduced UniRec, a novel generative recommendation system that fundamentally addresses the expressive gap between generative and discriminative approaches through a mechanism called Chain-of-Attribute (CoA). The core innovation lies in formalizing the problem via Bayes' theorem, showing that ranking by p(y|f,u) is equivalent to ranking by p(f|y,u), which can be factorized autoregressively over item features. This establishes that a generative model with full feature access is as expressive as its discriminative counterpart, with any practical gap stemming solely from incomplete feature coverage.
UniRec's CoA mechanism works by prefixing each Semantic ID (SID) sequence with structured attribute tokens—such as category, seller, and brand—before decoding the SID itself. This recovers the item-side feature crossing that discriminative models exploit. Because items sharing identical attributes cluster in adjacent SID regions, attribute conditioning yields measurable per-step entropy reduction, narrowing the search space and stabilizing beam search trajectories. The researchers also addressed deployment challenges including capacity-constrained SID with exposure-weighted capacity penalties to suppress token collapse, and Conditional Decoding Context (CDC) that injects scenario-conditioned signals at each decoding step.
The system employs a joint Reinforcement Fine-Tuning (RFT) and Direct Preference Optimization (DPO) framework to align the model with business objectives beyond distribution matching. Experimental results demonstrate substantial improvements, with UniRec outperforming the strongest baseline by +22.6% in HR@50 overall and +15.5% on high-value orders. Online A/B tests have confirmed significant gains across key business metrics, validating the practical effectiveness of bridging the generative-discriminative gap through structured attribute conditioning.
- UniRec uses Chain-of-Attribute to prefix Semantic IDs with structured attributes (category, seller, brand) before decoding, recovering item-side feature crossing
- Outperforms strongest baseline by +22.6% HR@50 overall and +15.5% on high-value orders in experiments
- Addresses deployment challenges with capacity-constrained SID penalties and Conditional Decoding Context for scenario-conditioned signals
Why It Matters
This breakthrough could revolutionize recommendation systems by combining the scalability of generative models with the precision of discriminative approaches.