Generative Recommender Systems: Redefining Recommendation Beyond Discriminative Modeling

Recommender system (RS) aims to capture user interest and recommend items that are potentially of the greatest interest to them based on their historical interactions. For decades, discriminative modeling has dominated the modeling paradigm, relying solely on ID-based collaborative data. However, practical challenges like data sparsity, diverse industrial scenarios, noisy behavioral sequences, and multimodal evolution limit its ability to generalize beyond observed interactions and capture user intent. In recent years, generative models like large language models (LLMs) have demonstrated strong semantic understanding, reasoning, and multimodal generation capabilities, offering a solution to complement discriminative models. However, their integration into RS faces difficulties of efficiency, compatibility, and task adaptation.

This dissertation advances a core thesis: recommender systems offer the potential to substantially mitigate the limitations of discriminative modeling (over-reliance on sparse data, poor generalization) by transitioning to a generative paradigm. Naturally, a core research question arises: how can traditional recommender systems adopting discriminative modeling be systematically redefined by a deployable generative paradigm? Meanwhile, three core challenges need to be systematically addressed: 1) Information Identification: What context-specific information is critical for improving recommendations? 2) Information Generation: How to effectively generate such information? 3) Information Integration: How to seamlessly integrate generated information into RS to boost performance?

To answer the core research question and tackle the challenges, we propose a three-stage progressive framework for building generative RS: 1) Lightweight Enhancement (prompt tuning, embedding fusion) to adapt generative capabilities to existing RS backbones. First, we propose (i) PLATE, a generative prompt-enhanced paradigm for multi-scenario recommendation and (ii) PAD, a generative LLM embedding-enhanced paradigm to for sequential recommendation.  2) Targeted Generation (data generation and knowledge reasoning) to address core limitations (data sparsity and knowledge scarcity). Specifically, we propose (i) Diff-MSR, a generative diffusion-based data augmentation paradigm for cold-start multi-scenario problem and (ii) LLM4MSR, a generative LLM-based knowledge reasoning-enhanced paradigm for multi-scenario recommendation. 3) Architecture Reshaping (LLM-as-backbone) to realize end-to-end generative recommendation. Specifically, we propose MME-SID, a generative sequential recommendation backbone leveraging multimodal semantic Identifiers and LLM. To sum up, these key contributions redefine recommender systems in the GenAI era, paving the way for next-generation recommender systems through generative paradigm.