[Paper Review] GRN: Generative Rerank Network for Context-wise Recommendation

Yufei Feng, Binbin Hu+4 more|arXiv (Cornell University)|Apr 2, 2021

Recommender Systems and Techniques32 references4 citations

TL;DR

GRN proposes a generative reranking framework that models contextual interactions in final ranking lists using a two-stage approach: an evaluator with Bi-LSTM and self-attention predicts item interaction probabilities, and a generator with GRU and pointer network learns context-aware reranking via policy gradient. The method achieves 5.2% and 6.1% improvements on PV and IPV in real-world deployment on Taobao.

ABSTRACT

Reranking is attracting incremental attention in the recommender systems, which rearranges the input ranking list into the final rank-ing list to better meet user demands. Most existing methods greedily rerank candidates through the rating scores from point-wise or list-wise models. Despite effectiveness, neglecting the mutual influence between each item and its contexts in the final ranking list often makes the greedy strategy based reranking methods sub-optimal. In this work, we propose a new context-wise reranking framework named Generative Rerank Network (GRN). Specifically, we first design the evaluator, which applies Bi-LSTM and self-attention mechanism to model the contextual information in the labeled final ranking list and predict the interaction probability of each item more precisely. Afterwards, we elaborate on the generator, equipped with GRU, attention mechanism and pointer network to select the item from the input ranking list step by step. Finally, we apply cross-entropy loss to train the evaluator and, subsequently, policy gradient to optimize the generator under the guidance of the evaluator. Empirical results show that GRN consistently and significantly outperforms state-of-the-art point-wise and list-wise methods. Moreover, GRN has achieved a performance improvement of 5.2% on PV and 6.1% on IPV metric after the successful deployment in one popular recommendation scenario of Taobao application.

Motivation & Objective

• To address the sub-optimal performance of greedy reranking methods that ignore contextual dependencies in final ranking lists.
• To model the mutual influence between items and their surrounding contexts in the final ranking list for more accurate interaction probability prediction.
• To learn a context-wise reranking strategy that optimizes for user engagement by considering sequential and relational item contexts.
• To enable effective deployment of context-aware reranking in low-latency, real-world recommendation systems like Taobao’s homepage.

Proposed method

• The evaluator uses Bi-LSTM to capture sequential dependencies in the final ranking list and self-attention to model mutual item interactions for refined interaction probability prediction.
• The generator employs GRU and an attention mechanism to select items step-by-step from the input list, using a pointer network to produce a permutation of the input list.
• Cross-entropy loss is used to train the evaluator, while policy gradient optimizes the generator based on the evaluator’s predicted rewards.
• The framework is trained in a two-stage manner: first, the evaluator is trained on labeled final lists; second, the generator is fine-tuned using policy gradient with rewards from the evaluator.
• The generator learns to produce contextually optimal rankings by distilling contextual knowledge from the trained evaluator.
• Efficient edge deployment is achieved via direct on-edge inference and an advance trigger mechanism to reduce latency in production.

Experimental results

Research questions

• RQ1 Can a context-aware reranking strategy outperform greedy reranking by modeling mutual item influences in the final list?
• RQ2 How effectively can a learned generator with policy gradient produce better final rankings than point-wise or list-wise models?
• RQ3 To what extent does incorporating contextual interaction probability improve recommendation performance in real-world industrial settings?
• RQ4 Can the proposed framework be efficiently deployed in low-latency production environments like Taobao’s homepage?

Key findings

• GRN consistently outperforms state-of-the-art point-wise and list-wise reranking methods on both benchmark and industrial datasets.
• On the industrial Rec dataset, GRN achieves a 5.2% improvement in PV and a 6.1% improvement in IPV over the deployed baseline PRM.
• Online A/B testing on Taobao’s Guess You Like scenario confirms GRN’s superiority, with significant gains in user engagement metrics.
• The generator’s context-aware strategy leads to more diverse and engaging recommendations, reducing user fatigue from repetitive categories.
• The framework achieves an average online inference latency of 32 milliseconds, with 30ms saved via edge deployment and request scheduling.
• The ablation study confirms that GRN G (with full context modeling) outperforms GRN G (-DR), proving the importance of comprehensive context integration.