Research map / Graph learning / Graph self-supervised learning

HGDL: Heterogeneous Graph Label Distribution Learning

NeurIPS 2024 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

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 HGDL, an end-to-end framework for label distribution learning on heterogeneous graphs that learns a meta-path topology via attention and a topology-and-feature consistency-aware graph transformer, trained with KL-divergence, on five datasets built with label distributions.

Problem

Label distribution learning has been studied only on IID data or homogeneous graphs; predicting label distributions for target nodes in heterogeneous graphs is open, complicated by node-type heterogeneity and inconsistency between topology and node features.

Approach

Proactively aggregates multiple meta-path graphs into one learned topology via node-level attention (before embedding learning), then harmonizes this topology with a feature-attention adjacency in a graph transformer block, optimizing KL-divergence plus an attention-diversity regularizer.

Key results

• HGDL achieves the lowest KL divergence on all five datasets, about 15% better than the second-best model on average (e.g., DRUG 0.392 vs 0.484 for GCNKL).
• On DBLP, HGDL reaches COD 0.019 and IND 0.943, versus 0.025 and 0.929 for HANKL, the strongest competitor.
• HGDL wins in 99 of 180 settings across metrics and datasets according to paired t-tests at 90% confidence.

Measured effects reported in the paper

Surprising findings

• On YELP, all methods perform nearly identically, which the authors attribute to the very low feature dimensionality (19).
• Clark Distance and KL divergence show a tradeoff in small-probability regions, so HGDL does not win on CLD in most settings.

Limitations

• No existing heterogeneous graph dataset has ground-truth label distributions, so the authors constructed five from existing graphs.
• Baselines are adapted heterogeneous GNNs with a KL loss because no prior method solves HGDL directly.
• Theoretical bounds rely on assumptions such as bounded features in an l2-ball, which the authors say are mild but are not verified.

Experimental setup

Datasets: DBLP ACM YELP DRUG URBAN

Benchmarks: DRUG ACM DBLP YELP URBAN

Metrics: Cosine Distance Canberra Distance Chebyshev Distance Clark Distance Intersection Score KL divergence

Baselines: GCNKL HANKL SeHGNNKL HGDL-noTH HGDL-noTransformer HGDL-ED

Builds on: HAN SeHGNN GAT Graph Transformer DropEdge

Closest papers by meaning

• GLDL: Graph Label Distribution Learning (AAAI 2024 · similarity 0.92)
• Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.88)
• NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (AAAI 2025 · similarity 0.87)
• HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.87)
• HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.87)
• Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective (NeurIPS 2024 · similarity 0.87)