Research map / Graph learning / GNN architectures

PANDA: Prior-guided Attentional Dual-path Architecture

NeurIPS 2026 · Graph learning › GNN architectures · approach: GNN generalization theory

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 PANDA, a plug-in for two-stage hypergraph message passing that mixes receiver-normalised attention with structural priors in a main path and adds a non-competitive gated auxiliary path, improving four HNN backbones by 5.18 points on average.

Problem

Hypergraph neural networks rely on fixed structural coefficients that propagate misleading information in heterophilic hypergraphs, or on receiver-normalised attention whose coefficients are competitive and cannot give several senders high weight simultaneously.

Approach

Interpolates per-incidence between learned receiver-normalised GATv2-style attention and the backbone's structural prior, and adds an auxiliary path whose gated weights are only locally bounded by the complement of the main coefficient, so they are not normalised across senders; directed roles are handled by factorising attention over head/tail sets.

Key results

• PANDA yields an average improvement of 5.18 percentage points over vanilla across four HNN backbones and eight datasets.
• On Telegram, PANDA improves PhenomNN by 31.01 points (44.81 to 75.82).
• Against a normalised auxiliary path (Norm Aux) with identical parameters, PANDA wins on 27 of 32 model-dataset pairs (+2.60 pp average).

Surprising findings

• On Citeseer and PubMed, PANDA slightly reduces accuracy for several backbones (PubMed PhenomNN -2.73), though changes are small.
• Disabling the auxiliary gates at test time drops Telegram accuracy by 54.0 points, indicating the trained model depends heavily on the auxiliary path.

Limitations

• Restricted to two-stage node-hyperedge-node message-passing HNNs.
• Reported gains on some homophilic datasets are negative or negligible; accuracy variance across runs is large on several datasets.
• Parameter counts and inference times vary inconsistently with PANDA, sometimes increasing overhead.

Experimental setup

Datasets: Telegram Citeseer PubMed Cornell Texas Squirrel Roman-empire email-Enron Wisconsin email-EU Questions

Benchmarks: Telegram Citeseer PubMed Roman-empire Cornell Texas Squirrel email-Enron Wisconsin email-EU Questions

Metrics: accuracy AUC-ROC

Baselines: HGNN HyperGCN PhenomNN GeDi-HNN GCN GAT HCHA

Builds on: GATv2 GeDi-HNN HGNN HyperGCN PhenomNN

Closest papers by meaning

• Training-Free Message Passing for Learning on Hypergraphs (ICLR 2025 · similarity 0.88)
• LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference (ICLR 2024 · similarity 0.87)
• When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · similarity 0.86)
• Deep Hypergraph Neural Networks with Tight Framelets (AAAI 2025 · similarity 0.86)
• Dissecting the Interplay of Attention Paths in a Statistical Mechanics Theory of Transformers (NeurIPS 2024 · similarity 0.86)
• SPHINX: Structural Prediction using Hypergraph Inference Network (ICML 2025 · similarity 0.86)