DiffKG: Knowledge Graph Diffusion Model for Recommendation (WSDM 2024)

Overview

DiffKG: Knowledge Graph Diffusion Model for Recommendation was published at WSDM 2024 (ACM International Conference on Web Search and Data Mining).

• Paper page: https://dl.acm.org/doi/abs/10.1145/3616855.3635850
• DOI: 10.1145/3616855.3635850
• Section: Recommender Engines

What the paper is about

Knowledge-graph-based recommenders typically exploit side information (entities and relations) to enrich user and item representations. However, real-world knowledge graphs are noisy and incomplete, which can propagate errors into the recommendation model.

DiffKG tackles this by bringing diffusion models to the knowledge graph recommendation setting:

• A forward diffusion process gradually corrupts the knowledge graph structure;
• A learned reverse process reconstructs/denoises the graph, producing more reliable triple structures;
• The denoised knowledge signals are integrated with collaborative filtering signals so that recommendation and graph generation are optimized jointly.

This positions DiffKG alongside other generative-modeling approaches to recommendation (VAE-based, score-based, and graph-diffusion methods), while specifically targeting the KG-recommendation coupling.

Where it fits in the literature

• KG-based recommendation: KGAT, RippleNet, KGCN and successors inject knowledge triples into CF models; DiffKG addresses the noise/incompleteness of these triples with a generative denoising view.
• Diffusion for graphs and RecSys: follows the broader trend of applying denoising diffusion to graph generation and to preference modeling in recommendation.
• Generative recommendation (Gen-Rec): part of the shift from discriminative ranking models toward generative formulations of the recommendation task.

Related entries in this collection

• 360Brew: A Decoder-only Foundation Model for Personalized Ranking
• Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
• Augmenting Netflix Search with In-Session Adapted Recommendations
• Bridging Language and Items for Retrieval and Recommendation
• Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
• EAGER-LLM: Enhancing Large Language Models as Recommenders

Note on details

This entry preserves the bibliographic record and positioning of the paper. For exact experimental datasets (e.g., Amazon/Last.fm-style benchmarks), backbone CF models, ablations, and reported nDCG/Recall numbers, please refer to the original PDF via the ACM link above. This post does not reproduce quantitative results to avoid misquoting values not verified against the full text.

References

• DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024. ACM DL