Non-autoregressive Generative Models for Reranking Recommendation

TL;DR

This work tackles the real-time reranking problem in multi-stage recommender systems by introducing NAR4Rec, a non-autoregressive generator within a generator-evaluator framework to efficiently explore permutation space $\mathcal{O}(A_{n}^{m})$. It introduces a matching model to handle dynamic candidate vocabularies, unlikelihood training to align with diverse user feedback, and contrastive decoding to model intra-list dependencies, complemented by a sequence evaluator to estimate listwise utility. Offline results on Avito and Kuaishou show that NAR4Rec outperforms strong baselines in AUC, Recall, and LogLoss, while online A/B tests reveal meaningful gains in views, watch time, and positive interactions. The framework delivers substantial speedups over autoregressive baselines and has been deployed in a real-world app with hundreds of millions of daily users, underscoring its practical impact for scalable, high-quality recommendations.

Abstract

Contemporary recommendation systems are designed to meet users' needs by delivering tailored lists of items that align with their specific demands or interests. In a multi-stage recommendation system, reranking plays a crucial role by modeling the intra-list correlations among items. The key challenge of reranking lies in the exploration of optimal sequences within the combinatorial space of permutations. Recent research proposes a generator-evaluator learning paradigm, where the generator generates multiple feasible sequences and the evaluator picks out the best sequence based on the estimated listwise score. The generator is of vital importance, and generative models are well-suited for the generator function. Current generative models employ an autoregressive strategy for sequence generation. However, deploying autoregressive models in real-time industrial systems is challenging. To address these issues, we propose a Non-AutoRegressive generative model for reranking Recommendation (NAR4Rec) designed to enhance efficiency and effectiveness. To tackle challenges such as sparse training samples and dynamic candidates, we introduce a matching model. Considering the diverse nature of user feedback, we employ a sequence-level unlikelihood training objective to differentiate feasible sequences from unfeasible ones. Additionally, to overcome the lack of dependency modeling in non-autoregressive models regarding target items, we introduce contrastive decoding to capture correlations among these items. Extensive offline experiments validate the superior performance of NAR4Rec over state-of-the-art reranking methods. Online A/B tests reveal that NAR4Rec significantly enhances the user experience. Furthermore, NAR4Rec has been fully deployed in a popular video app Kuaishou with over 300 million daily active users.

Figures (4)

• Figure 1: Comparison between autoregressive model (left) and non-autoregressive model (right). Auto-regressive models generate items sequentially while non-autoregressive models generate all items simultaneously.
• Figure 2: Architectural Overview of the Generator and Evaluator Models. The evaluator evaluates multiple sequences generated by the generator and estimates listwise score to select the optimal sequence.
• Figure 3: The comparison between NAR4Rec and PIER on Avito with different learning rate, batch size and epoch.
• Figure 4: The effect of parameter weight $\omega$ and margin $\rho$ in contrastive loss and the $\alpha$ in contrastive decoding.