Research map / Graph learning / GNN architectures

NoiseHGNN: Synthesized Similarity Graph-Based Neural Network for Noised Heterogeneous Graph Representation Learning

AAAI 2025 · Graph learning › GNN architectures · approach: Graph structure learning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NoiseHGNN, which builds a feature-similarity graph from node features and uses a shared-parameter similarity-aware attention encoder plus metapath-target contrastive loss to train heterogeneous GNNs robustly under 30% erroneous links.

Problem

Heterogeneous graph representation learning degrades when the graph structure contains noisy (erroneous) links, and the homogeneity assumption used to correct noisy homogeneous graphs via similarity graphs fails because similar nodes in heterogeneous graphs are not directly linked.

Approach

Synthesize a kNN similarity graph from original node features, encode both the noisy graph and similarity graph with a shared-parameter similarity-aware attention encoder, supervise both with labels, and contrast a metapath-based graph from the noisy graph with a target-based graph from the similarity graph via a scaled cosine loss.

Key results

• On 30% noised graphs, NoiseHGNN reaches 71.96 Macro-F1 on DBLP and 60.46 Macro-F1 on IMDB, versus 68.23 and 58.21 for the best prior methods (SlotGAT, simpleHGN).
• On PubMed_NC (30% noise), NoiseHGNN reports 50.92 Macro-F1 and 55.81 Micro-F1, above SlotGAT's 44.76 and 50.88.
• On clean original datasets NoiseHGNN is not SOTA (e.g., ACM 91.95 vs Seq-HGNN 94.41 Macro-F1 on DBLP-clean row table), and link prediction results are comparable rather than leading.

Measured effects reported in the paper

Surprising findings

• On original clean datasets the method does not achieve state-of-the-art, e.g., Seq-HGNN and SlotGAT outperform it on DBLP and IMDB.
• Ablation shows removing the graph synthesizer causes a large drop on IMDB (Macro-F1 43.46 vs 60.46).

Limitations

• Does not reach state-of-the-art on clean (non-noised) heterogeneous graph benchmarks.
• Similarity graph and kNN sparsification add cost; large-scale variant uses batch-local approximation.
• Evaluated only on node classification with simulated random link noise at fixed rates.

Experimental setup

Datasets: DBLP IMDB ACM Freebase PubMed_NC LastFM PubMed_LP

Benchmarks: DBLP IMDB ACM Freebase PubMed_NC LastFM PubMed_LP

Metrics: Macro-F1 Micro-F1 ROC-AUC MRR training time inference time peak GPU memory silhouette score

Baselines: GCN GAT RGCN RSHN HAN HetGNN MAGNN HGT simpleHGN Space4HGNN Seq-HGNN SlotGAT

Builds on: GAT HGSL SimpleHGN SlotGAT

Closest papers by meaning

• HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.88)
• Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.88)
• Training Robust Graph Neural Networks by Modeling Noise Dependencies (NeurIPS 2025 · similarity 0.87)
• HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.87)
• NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise (NeurIPS 2024 · similarity 0.87)
• Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.87)