EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (KDD 2024)

This post indexes the KDD 2024 paper EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration, a two-stream generative recommender (Gen-Rec) model for sequential recommendation.

• Publication: KDD 2024
• Link: https://dl.acm.org/doi/abs/10.1145/3637528.3671775
• Category: Sequential Recommendation

Context and Motivation

Large-scale recommendation systems have long faced challenges in efficiency, scalability, and user intent understanding. Traditional pipelines treat retrieval, ranking, and generation as separate stages. In the LLM era, generative recommenders aim to unify these stages, but a core tension remains: collaborative signals from user behavior are sparse, item catalogs are huge, and business objectives involve multi-dimensional trade-offs.

EAGER addresses this by proposing a two-stream architecture that combines:

• Behavioral stream: collaborative filtering signals learned from user interaction sequences
• Semantic stream: content/semantic understanding of items

with a behavior-semantic collaboration mechanism that lets the two representations inform each other within a generative recommendation framework.

Position in the Field

The post situates EAGER within the evolution of sequential recommendation:

1. Matrix factorization → deep CTR models → sequential Transformers (e.g., BERT4Rec) 2. Generative recommendation (Gen-Rec), where item indexing and generation replace discriminative ranking 3. LLM-based recommenders offering semantic priors and cold-start ability, at the cost of inference latency and hallucination risk

Key Takeaways for Practitioners

• Architecture: cascade retrieval + rerank + generation remains mainstream; generative approaches reduce cascade errors but face index-update challenges
• Data: high-quality interaction logs matter as much as model design; synthetic data must avoid leakage and distribution shift
• Evaluation: offline metrics (nDCG, Recall@k, Hit@k) increasingly diverge from online satisfaction; combine with human or LLM-as-judge evaluation
• Deployment: latency, cost, interpretability, and safety are hard constraints — optimize beyond academic benchmarks

Related Entries

• BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
• Efficient On-Device Session-Based Recommendation (TOIS)
• How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
• LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with LLMs
• Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models
• Multi-Behavior Sequential Transformer Recommender (SIGIR 2024)

> Note: Specific numerical results and full architecture details should be verified against the original paper PDF, as this entry is based on the abstract and public metadata.

References

• Original paper: EAGER, KDD 2024