A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys) - KDD 2024
Overview
- Title: A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys), KDD 2024
- Source: https://dl.acm.org/doi/abs/10.1145/3637528.3671474
- Type: Survey
- Section: Recommender Engine Surveys
- Unified perspective: The survey consolidates previously scattered generative-recommendation work into a comparable framework, decomposing systems into representation learning, retrievers, rerankers, planners, generators, and feedback mechanisms.
- Motivation: Traditional pipelines split retrieval, ranking, and generation, which struggles to meet LLM-era demands for natural language interaction, multi-hop reasoning, and up-to-date knowledge in recommendation.
- Taxonomy highlights:
- *Modeling paradigm*: discriminative retrieval (two-tower, cross-encoders) vs. generative retrieval (DSI-style docid generation); the former is mature and scalable, the latter simplifies the cascade but faces index-update costs.
- *LLM integration*: RAG, agentic search, and tool use offer flexibility and interpretability at the cost of latency and error propagation.
- *Objectives*: relevance, diversity, and freshness balanced via multi-objective LTR, RLHF, and online learning.
- *Evaluation*: offline metrics (nDCG@10, MRR, Recall@k), human preference, LLM-as-judge, and online A/B testing; gaps between offline metrics and real user satisfaction are widening.
- 2019–2021: BERT rerankers and DPR establish neural retrieval foundations.
- 2022–2023: RAG and retrieval-generation fusion become mainstream.
- 2024: Conversational/agentic search and Gen-RecSys research surge.
- 2025 onward: RL-trained search agents, deep research systems, and GraphRAG emerge as growth areas.
- Lack of unified, reproducible benchmarks; private data limits reproducibility.
- LLM-judge evaluation bias requires cross-validation with human assessment.
- Hallucination, safety, latency, and cost constraints in production systems.
- Cross-lingual, multimodal, and fairness-aware extension for generative recommenders.
- A Comprehensive Survey on Cross-Domain Recommendation (arXiv:2503.14110)
- A Comprehensive Survey on Retrieval Methods in Recommender Systems (arXiv:2407.21022)
- A Survey of Generative Search and Recommendation in the Era of Large Language Models (arXiv:2404.16924)
- A Survey on LLM-powered Agents for Recommender Systems (arXiv:2502.10050)
- Original paper: A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys), KDD 2024. ACM DL
This KDD 2024 survey systematically reviews generative recommender systems (Gen-RecSys), covering how generative models and large language models (LLMs) reshape the retrieval, ranking, and generation stages of modern recommendation pipelines.
Key Points
Timeline of the Field
Open Problems
Practical Takeaways
1. Architecture: Cascade retrieval + reranking + generation remains mainstream, but agentic paradigms make retrieval strategies themselves learnable. 2. Data: High-quality instruction data and session logs matter; synthetic data risks knowledge leakage and distribution shift. 3. Deployment: Latency budgets, token cost, explainability, and safety filtering are hard constraints that pure academic benchmarks overlook.