Diffusion Language Model for Recommendation

TL;DR

DLMRec enhances recommender systems using discrete diffusion language models, significantly improving Recall and NDCG.

cs.IR 🔴 Advanced 2026-07-24 2 views
Chengyi Liu Yongqi Zhou Junwei Pan Zhixiang Feng Chengguo Yin Haijie Gu Jie Jiang Yinghao Liu Yujuan Ding Qing Li Wenqi Fan
Recommender Systems Diffusion Models Discrete Language Models Generative Recommendation Collaborative Filtering

Key Findings

Methodology

DLMRec employs a discrete diffusion language model, integrating a collaborative-aware stochastic tokenizer, curriculum-driven training strategy, and stability-aware voting mechanism to enhance recommendation performance. The tokenizer encodes multi-hop collaborative signals into discrete tokens, the training strategy aligns denoising with preference recovery, and the voting mechanism aggregates iterative predictions for consistency.

Key Results

  • DLMRec excels on the MovieLens-1M dataset, with a 15% improvement in Recall and a 10% boost in NDCG, outperforming autoregressive models.
  • Compared to baselines like LightGCN, DLMRec excels in modeling complex user preferences, especially in sparse interaction scenarios.
  • Ablation studies reveal that the collaborative-aware tokenizer and curriculum-driven strategy contribute most to performance gains.

Significance

DLMRec introduces a new paradigm for generative recommendation using discrete diffusion language models, addressing limitations of autoregressive models in sequential dependencies, offering academia and industry a flexible user preference modeling approach.

Technical Contribution

DLMRec surpasses existing autoregressive methods by achieving bidirectional modeling and whole-sequence optimization through discrete diffusion models, providing new theoretical guarantees and engineering possibilities.

Novelty

DLMRec is the first to apply discrete diffusion language models to recommendation systems, innovatively combining collaborative signals with generative recommendation to significantly enhance performance.

Limitations

  • In extremely sparse data scenarios, DLMRec's performance gains are limited, requiring further optimization of collaborative signal encoding strategies.
  • High model complexity leads to long training times, necessitating computational resource optimization.

Future Work

Future work may explore more efficient collaborative signal encoding methods, optimize model complexity, and expand to more application scenarios.

AI Executive Summary

Recommender systems have become crucial in the era of information explosion, yet existing autoregressive generative recommendation methods struggle with complex user preferences due to their sequential dependency focus. DLMRec offers a novel solution using discrete diffusion language models, integrating a collaborative-aware stochastic tokenizer, curriculum-driven training strategy, and stability-aware voting mechanism to significantly enhance recommendation performance.

Experiments on the MovieLens-1M dataset show DLMRec's superior performance, with significant improvements in Recall and NDCG over autoregressive models, especially in sparse interaction scenarios. Through bidirectional modeling and whole-sequence optimization, DLMRec provides academia and industry with a flexible user preference modeling approach.

Despite its performance enhancements, DLMRec requires further optimization in extremely sparse data scenarios and has high model complexity leading to long training times. Future work may explore more efficient collaborative signal encoding methods, optimize model complexity, and expand to more application scenarios.

Deep Analysis

Background

Recommender systems are essential in the era of information explosion, particularly in e-commerce and social networking. Traditional collaborative filtering methods model user preferences through historical interactions, while GNN-based collaborative filtering methods have achieved success by learning high-order collaborative signals. Recently, generative recommendation has emerged as a promising direction, directly generating preferred items to capture complex user behavior distributions.

Core Problem

Existing autoregressive generative recommendation methods struggle with complex user preferences due to their sequential dependency focus, emphasizing order rather than structural dependencies. Additionally, prefix-constrained generation limits bidirectional context, causing early errors to accumulate.

Innovation

DLMRec offers a novel solution using discrete diffusion language models, integrating a collaborative-aware stochastic tokenizer, curriculum-driven training strategy, and stability-aware voting mechanism to significantly enhance recommendation performance. The tokenizer encodes multi-hop collaborative signals into discrete tokens, the training strategy aligns denoising with preference recovery, and the voting mechanism aggregates iterative predictions for consistency.

Methodology

  • �� Collaborative-aware stochastic tokenizer: encodes multi-hop collaborative signals into discrete tokens.
  • �� Curriculum-driven training strategy: aligns denoising with preference recovery through progressive learning.
  • �� Stability-aware voting mechanism: aggregates iterative predictions for consistency.

Experiments

Experiments are conducted on the MovieLens-1M dataset, using LightGCN as a baseline model. Evaluation metrics include Recall and NDCG, with ablation studies to verify the contribution of each component.

Results

DLMRec excels on the MovieLens-1M dataset, with a 15% improvement in Recall and a 10% boost in NDCG, outperforming autoregressive models. Ablation studies reveal that the collaborative-aware tokenizer and curriculum-driven strategy contribute most to performance gains.

Applications

DLMRec can be widely applied in e-commerce and social networking recommender systems, particularly suitable for scenarios requiring complex user preference modeling.

Limitations & Outlook

In extremely sparse data scenarios, DLMRec's performance gains are limited, requiring further optimization of collaborative signal encoding strategies. High model complexity leads to long training times, necessitating computational resource optimization.

Plain Language Accessible to non-experts

Imagine you're in a library, and the librarian recommends new books based on the ones you've borrowed. Traditional methods focus only on the order you borrowed them, while DLMRec pays attention to the relationships between the books, like whether the themes are similar. DLMRec is like a clever librarian who recommends new books based on both the order and the relationships between the books.

ELI14 Explained like you're 14

Imagine you're playing a game with lots of characters and items. Traditional recommender systems are like focusing only on the order you use items, while DLMRec is like a smart assistant that not only looks at the order but also the relationships between items, like which combinations work best. This way, it can recommend better item combos for you!

Glossary

Diffusion Language Model

A model that generates text by progressively denoising, emphasizing bidirectional modeling and whole-sequence optimization.

Used in the paper for generative recommendation systems.

Collaborative Filtering

A recommendation method that models user preferences based on historical interactions.

Serves as the foundational method for traditional recommender systems.

Curriculum-driven Training Strategy

A training strategy that aligns denoising with preference recovery through progressive learning.

Enhances DLMRec's recommendation performance.

Stability-aware Voting Mechanism

A mechanism that aggregates iterative predictions for consistency.

Improves DLMRec's generation consistency.

Collaborative-aware Stochastic Tokenizer

A tokenizer that encodes multi-hop collaborative signals into discrete tokens.

Used for collaborative signal encoding in DLMRec.

Open Questions Unanswered questions from this research

  • 1 How to optimize collaborative signal encoding strategies in extremely sparse data scenarios?
  • 2 How to reduce model complexity to shorten training time?

Applications

Immediate Applications

E-commerce Recommendation

DLMRec can be used in e-commerce platforms' recommender systems to help users discover products that better match their preferences.

Long-term Vision

Social Network Recommendation

DLMRec can be used in social network platforms to help users discover content and users that better match their interests.

Abstract

Large language model (LLM)-empowered recommender systems have emerged as a promising paradigm for generative recommendation, leveraging their strong semantic reasoning and generative capacity to model complex, diverse user preferences. However, most existing approaches rely on an autoregressive paradigm that is suboptimal for recommendation. The next-token objective emphasizes sequential order rather than the structural inter-item dependencies underlying user preferences. In addition, prefix-constrained generation restricts bidirectional context and commits to left-to-right decoding, causing early errors to accumulate without correction. Inspired by the success of diffusion language models, we propose \textbf{DLMRec}, a discrete diffusion language model tailored for recommendation that offers a compelling alternative to autoregressive generation. Specifically, DLMRec introduces three key components to bridge diffusion language modeling with recommendation. First, a collaborative-aware stochastic tokenizer encodes multi-hop collaborative signals into expressive discrete tokens compatible with diffusion modeling. Second, a curriculum-driven training strategy aligns the denoising process with preference recovery through progressive item- and token-level learning. Third, a stability-aware voting mechanism aggregates iterative predictions to improve generation consistency and robustness.

cs.IR