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

Source: arXiv - 2603.16800v1

Overview

The paper introduces RaDAR, a new framework that blends graph neural networks (GNNs) with graph contrastive learning (GCL) to boost collaborative‑filtering recommendation systems. By tackling two common pain points—noisy edge perturbations and extreme data sparsity—RaDAR delivers more reliable recommendations, especially in real‑world settings where user‑item interaction graphs are messy and incomplete.

Key Contributions

• Asymmetric contrastive learning with global negative sampling – keeps the semantic meaning of positive pairs while aggressively pushing apart unrelated items, reducing the “semantic drift” caused by random augmentations.
• Diffusion‑guided augmentation – injects noise progressively through a diffusion process and then denoises it, yielding robust graph views that preserve global structure.
• Relation‑aware edge refinement – a lightweight denoising module that re‑weights edges on‑the‑fly based on latent node semantics, effectively pruning spurious connections.
• Unified two‑view generation pipeline – combines a graph generative model (global view) with the relation‑aware denoiser (local view) to capture both macro‑level patterns and micro‑level relational nuances.
• Extensive empirical validation – outperforms 12 state‑of‑the‑art baselines on three public recommendation benchmarks, with the biggest gains observed under high noise and high sparsity conditions.

Methodology

1. Graph Construction – The user‑item interaction matrix is turned into a bipartite graph where nodes are users or items and edges represent observed interactions.
2. View Generation
   • Global View: A diffusion‑based generative model adds controlled stochastic perturbations to the adjacency matrix, then runs a diffusion‑guided denoising step. This preserves long‑range connectivity while still exposing the model to varied structures.
   • Local (Relation‑Aware) View: A relation‑aware denoising network examines each edge’s latent embeddings (produced by a shallow GNN) and adjusts its weight, effectively filtering out noisy or irrelevant connections.
3. Asymmetric Contrastive Objective – The two views are fed into a shared encoder. Positive pairs are formed by the same node across the two views, while negatives are sampled globally (any other node in the batch), not just from the immediate neighbourhood. The loss encourages the embeddings of the same node to align while pushing unrelated nodes apart.
4. Training Loop – The contrastive loss is combined with a standard recommendation loss (e.g., Bayesian Personalized Ranking) so the model learns both to distinguish good from bad recommendations and to maintain robust graph representations.

The whole pipeline is end‑to‑end differentiable and can be dropped into existing GNN‑based recommenders with minimal code changes.

Results & Findings

• Noise robustness: When synthetic random edges (up to 30 % of total edges) are added, RaDAR’s HR@10 drops <2 % while competing methods lose >10 %.
• Sparsity resilience: With only 10 % of interactions retained, RaDAR still outperforms baselines by 7–9 % absolute HR@10.
• Ablation studies show that removing either diffusion augmentation or relation‑aware refinement reduces performance by ~3 % each, confirming that both components are complementary.

Overall, the experiments demonstrate that RaDAR can maintain high recommendation quality even when the underlying graph is noisy or extremely sparse.

Practical Implications

• Cleaner embeddings with less data: Developers can achieve strong recommendation performance without needing massive interaction logs, which is valuable for new platforms or niche domains.
• Plug‑and‑play module: The asymmetric contrastive loss and diffusion‑based augmentor are framework‑agnostic (PyTorch, TensorFlow, DGL, etc.), making it straightforward to integrate into existing pipelines.
• Robustness to data‑drift: In production, user behaviour changes and occasional logging errors introduce noise; RaDAR’s denoising view helps the model stay stable over time, reducing the frequency of costly retraining.
• Edge‑level interpretability: The relation‑aware edge weights can be inspected to identify suspicious or low‑quality interactions, offering a diagnostic tool for data engineers.
• Scalable training: The global negative sampling strategy avoids the quadratic cost of pairwise negatives, keeping GPU memory usage comparable to standard BPR training.

Limitations & Future Work

• Computational overhead: The diffusion‑guided augmentation adds extra forward passes, increasing training time by ~15–20 % compared to vanilla GNN recommenders.
• Hyper‑parameter sensitivity: The noise schedule for diffusion and the threshold for edge re‑weighting require careful tuning on each dataset.
• Cold‑start users/items: While RaDAR mitigates sparsity, it still relies on at least a few interactions; extending the framework with side‑information (e.g., textual or visual features) is an open direction.
• Theoretical analysis: The paper provides empirical evidence for robustness but lacks a formal guarantee on how diffusion noise bounds affect representation quality. Future work could explore tighter theoretical links between diffusion processes and contrastive alignment.

Authors

• Yixuan Huang
• Jiawei Chen
• Shengfan Zhang
• Zongsheng Cao

Paper Information

• arXiv ID: 2603.16800v1
• Categories: cs.LG
• Published: March 17, 2026
• PDF: Download PDF