RaDAR: Relation-aware Diffusion-Asymmetric Graph Contrastive Learning for Recommendation

Yixuan Huang; Jiawei Chen; Shengfan Zhang; Zongsheng Cao

cs.LGMar 17, 2026

RaDAR: Relation-aware Diffusion-Asymmetric Graph Contrastive Learning for Recommendation

Authors: Yixuan Huang, Jiawei Chen, Shengfan Zhang, Zongsheng Cao

Organizations: National University of Defense Technology, Changsha, China · Ant Group, Chongqing, China · Tsinghua University, Beijing, China

Abstract

Collaborative filtering (CF) recommendation has been significantly advanced by integrating Graph Neural Networks (GNNs) and Graph Contrastive Learning (GCL). However, (i) random edge perturbations often distort critical structural signals and degrade semantic consistency across augmented views, and (ii) data sparsity hampers the propagation of collaborative signals, limiting generalization. To tackle these challenges, we propose RaDAR (Relation-aware Diffusion-Asymmetric Graph Contrastive Learning Framework for Recommendation Systems), a novel framework that combines two complementary view generation mechanisms: a graph generative model to capture global structure and a relation-aware denoising model to refine noisy edges. RaDAR introduces three key innovations: (1) asymmetric contrastive learning with global negative sampling to maintain semantic alignment while suppressing noise; (2) diffusion-guided augmentation, which employs progressive noise injection and denoising for enhanced robustness; and (3) relation-aware edge refinement, dynamically adjusting edge weights based on latent node semantics. Extensive experiments on three public benchmarks demonstrate that RaDAR consistently outperforms state-of-the-art methods, particularly under noisy and sparse conditions.

Figures & tables

Figure 1. ACL and diffusion model on a user–item graph, illustrating how standard diffusion misses two-hop monophily where indirectly connected users share similar preferences.

Figure 2. RaDAR framework architecture: The left section shows two view generators extracting complementary graph representations. The right section demonstrates the contrastive learning process with diffusion model-based graph generation and joint optimization through InfoNCE, IB, and BPR losses.

Phase	Objective	Params
1	$L_{\text{bpr}},L_{\text{diff-ssl}},\\|\Theta\\|_{F}^{2}$	User-item embeds
2	$\mathcal{L}_{IB}$ (Info. bottleneck)	User-item embeds
3	$\mathcal{L}_{gen}+\mathcal{L}_{den}$	View generators

Table 1. Training phases in our framework.

Dataset	Users	Items	Links	Interaction Types
Tmall	31,882	31,232	1,451,29	View, Favorite, Cart, Purchase
RetailRocket	2,174	30,113	97,381	View, Cart, Transaction
IJCAI15	17,435	35,920	799,368	View, Favorite, Cart, Purchase

Table 2. Statistics of weighted-edge datasets.

Dataset	Users	Items	Interactions	Density
Last.FM	1,892	17,632	92,834	$2.8\times 10^{-3}$
Yelp	42,712	26,822	182,357	$1.6\times 10^{-4}$
BeerAdvocate	10,456	13,845	1,381,094	$9.5\times 10^{-3}$

Table 3. Statistics of the experimental datasets.

Model	Last.FM				Yelp				BeerAdvocate
Model	Recall@20	NDCG@20	Recall@40	NDCG@40	Recall@20	NDCG@20	Recall@40	NDCG@40	Recall@20	NDCG@20	Recall@40	NDCG@40
BiasMF	0.1879	0.1362	0.2660	0.1653	0.0532	0.0264	0.0802	0.0321	0.0996	0.0856	0.1602	0.1016
NCF	0.1130	0.0795	0.1693	0.0952	0.0304	0.0143	0.0487	0.0187	0.0729	0.0654	0.1203	0.0754
AutoR	0.1518	0.1114	0.2174	0.1336	0.0491	0.0222	0.0692	0.0268	0.0816	0.0650	0.1325	0.0794
PinSage	0.1690	0.1228	0.2402	0.1472	0.0510	0.0245	0.0743	0.0315	0.0930	0.0816	0.1553	0.0980
STGCN	0.2067	0.1558	0.2940	0.1821	0.0562	0.0282	0.0856	0.0355	0.1003	0.0852	0.1650	0.1031

Table 4. Performance Metrics for Various Models

Model	Tmall		RetailRocket		IJCAI15
Model	R@20	N@20	R@20	N@20	R@20	N@20
BPR	0.0248	0.0131	0.0308	0.0237	0.0051	0.0037
PinSage	0.0368	0.0156	0.0423	0.0248	0.0101	0.0041
NGCF	0.0399	0.0169	0.0405	0.0257	0.0091	0.0035
NMTR	0.0441	0.0192	0.0460	0.0265	0.0108	0.0048
MBGCN	0.0419	0.0179	0.0492	0.0258	0.0112	0.0045

Table 5. Weighted-edge results under DiffGraph settings (Recall@20 / NDCG@20).

Model	Variant	Last.FM		Yelp		Beer
Model	Description	Recall	NDCG	Recall	NDCG	Recall	NDCG
Baseline	SOTA SSL	0.2603	0.1911	0.0873	0.0439	0.1216	0.1015
RaDAR	Gen+Gen	0.2665	0.1936	0.0900	0.0456	0.1226	0.1027
	Gen+Linear	0.2698	0.1986	0.0910	0.0461	0.1247	0.1050
	w/o D-ACL	0.2652	0.1934	0.0904	0.0458	0.1250	0.1036
	w/ ACL only	0.2720	0.1962	0.0911	0.0461	0.1264	0.1057

Table 6. Ablation study on key components of RaDAR.

Panel of Figure 3 — Figure 3. Impact of Noise Ratio (5%–25%) on Performance Degradation

Panel of Figure 4 — Figure 4. Performance variation with ACL ratio $\lambda$ . Last.FM peaks Recall@20 at $\lambda=5.5$ , NDCG@20 at $\lambda=3.5$ . Yelp peaks Recall@20 at $\lambda=1.5$ , NDCG@20 at $\lambda=1.0$ . Higher $\lambda$ values enhance relation-aware denoising for Last.FM, while Yelp requires balanced contributions due to interaction sparsity.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5. Performance analysis across five user and item interaction sparsity levels on Yelp dataset.

Graph Contrastive Learning

Graph contrastive learning trains graph representations by aligning related nodes or graph views while separating structurally or semantically different examples.

See all 36 papers

Adaptive Item-based Collaborative Structures via Noise Rescheduling in Diffusion for Generative Recommendation
Aug 24, 2026Jiaqi Wang, Tianying Liu, Heng Chang +3Generative Recommendation Diffusion Models
Fudan University, Shanghai, China · Tongji University, Shanghai, China
Bridging the Topological Divide: Dual-Branch Graph Distillation with Semantic Priors for Robust Cold-Start Recommendation
Oct 14, 2026Yuanfeng QiuCold-Start Recommendation Collaborative Filtering
The Hong Kong University of Science and Technology (Guangzhou)
Multi-Level Graph Attention Network Contrastive Learning for Knowledge-Aware Recommendation
May 8, 2026Zhifei Hu, Feng XiaGraph Attention Networks Multi-View Learning
Durham University, Durham, United Kingdom · University of Sheffield, Sheffield, United Kingdom

Aug 24, 2026J/K move · Enter open · S save

Adaptive Item-based Collaborative Structures via Noise Rescheduling in Diffusion for Generative Recommendation

Jiaqi Wang, Tianying Liu, Heng Chang, Jihong Guan+2

Fudan University, Shanghai, China · Tongji University, Shanghai, China

Generative Recommendation Diffusion Models Collaborative Filtering

Open paper AskPDF arXiv

Discrete Diffusion Models (DDMs) have recently been introduced to recommendation systems, modeling user history as a token generation process via iterative denoising. However, while effective at capturing user-level sequential patterns, these methods often fail to explicitly integrate item-based collaborative filtering information, a critical component for accurate recommendation. This deficiency manifests in two key aspects: (1) the item representation is often semantic-focused, lacking collaborative priors for diffusion training; and (2) the denoising process employs a uniform noise schedule, treating all tokens indiscriminately and ignoring item-level adaptive structural dependencies. To bridge this gap, we propose ANR-DiffRec, a unified framework designed to encode item-based collaborative structures into discrete diffusion for generative recommendation. First, we explicitly incorporate an item co-occurrence matrix to guide semantic ID generation, providing a structured collaborative prior for discrete diffusion training. Second, we introduce an item-based adaptive noise rescheduling mechanism that dynamically adjusts denoising weights according to both local contextual recoverability and behavior-aware item dependencies. Specifically, the proposed strategy jointly models intra-item structural context and inter-item collaborative signals, enabling structure-aware denoising during diffusion training. Extensive experiments on multiple benchmarks demonstrate that our method consistently outperforms state-of-the-art generative recommendation models. Code: https://github.com/CalmaQi/ANR-DiffRec.

Abstract

Graph Contrastive Learning

Explore similar work

Adaptive Item-based Collaborative Structures via Noise Rescheduling in Diffusion for Generative Recommendation

Bridging the Topological Divide: Dual-Branch Graph Distillation with Semantic Priors for Robust Cold-Start Recommendation

Multi-Level Graph Attention Network Contrastive Learning for Knowledge-Aware Recommendation