🕸️ Graph Learning¶

🤖 AAAI2026 · 38 paper notes

Adaptive Initial Residual Connections for GNNs with Theoretical Guarantees: This paper proposes Adaptive Initial Residual Connections (Adaptive IRC), which allows each node to have a personalized residual strength learned from its initial features. It provides the first theoretical proof of a positive lower bound on the Dirichlet energy of initial residual connections with activation functions (guaranteeing the absence of over-smoothing), and introduces a PageRank-based heuristic variant that achieves comparable or superior performance without learning additional parameters.
Adaptive Riemannian Graph Neural Networks: This paper proposes ARGNN, a framework that learns a continuous, anisotropic diagonal Riemannian metric tensor for each node in a graph, enabling adaptive capture of local geometric properties across different graph regions (hierarchical structures vs. dense communities). ARGNN unifies and outperforms geometric GNN methods based on fixed curvature or discrete mixed-curvature spaces.
Are Graph Transformers Necessary? Efficient Long-Range Message Passing with Fractal Nodes in MPNNs: This paper proposes Fractal Nodes (FN) to enhance long-range message passing in MPNNs. Subgraph-level aggregation nodes are generated via METIS graph partitioning, combined with low-pass and high-pass filters (LPF+HPF) and a learnable frequency parameter \(\omega\). MLP-Mixer is adopted for cross-subgraph communication. The approach achieves \(O(L(|V|+|E|))\) linear complexity while matching or surpassing Graph Transformer performance, earning an AAAI Oral.
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation: This paper proposes ARG-Designer, which reformulates multi-agent system topology design as a conditional autoregressive graph generation task. Rather than pruning from template graphs, the model incrementally generates agent nodes and communication edges from scratch. ARG-Designer achieves state-of-the-art performance across 6 benchmarks (average 92.78%), reduces token consumption by approximately 50% compared to G-Designer, and supports role expansion without retraining.
Assessing LLMs for Serendipity Discovery in Knowledge Graphs: A Case for Drug Repurposing: This paper proposes SerenQA, the first framework to formally define the serendipity discovery task in knowledge graph question answering. It introduces an information-theoretic RNS metric, an expert-annotated drug repurposing benchmark dataset, and a three-stage LLM evaluation pipeline. The work reveals that current LLMs perform reasonably on retrieval tasks but have substantial room for improvement in serendipitous exploration.
Beyond Fixed Depth: Adaptive Graph Neural Networks for Node Classification Under Varying Homophily: This paper proposes AD-GNN, which theoretically analyzes node-level homophily/heterophily characteristics and adaptively assigns different aggregation depths to individual nodes, enabling unified handling of node classification on both homophilic and heterophilic graphs within a single framework.
BugSweeper: Function-Level Detection of Smart Contract Vulnerabilities Using Graph Neural Networks: This paper proposes BugSweeper, which constructs function-level abstract syntax graphs (FLAG) and designs a two-stage GNN architecture to enable end-to-end smart contract vulnerability detection without expert-defined rules, achieving an F1 of 98.57% on reentrancy attack detection.
Commonality in Few: Few-Shot Multimodal Anomaly Detection via Hypergraph-Enhanced Memory: This paper proposes CIF, which leverages hypergraphs to extract intra-class structural commonalities from a small number of training samples, guiding memory bank construction and retrieval for few-shot multimodal industrial anomaly detection, achieving state-of-the-art performance.
EchoLess: Label-Based Pre-Computation for Memory-Efficient Heterogeneous Graph Learning: Echoless-LP eliminates training label leakage (the echo effect) caused by multi-hop message passing in label pre-computation via Partition-Focused Echoless Propagation (PFEP). Combined with an Asymmetric Partition Scheme (APS) and a PostAdjust mechanism to address information loss and distribution shift introduced by partitioning, the method remains memory-efficient, is compatible with arbitrary message-passing operators, and achieves state-of-the-art performance on multiple heterogeneous graph benchmarks.
Enhancing Logical Expressiveness in GNNs via Path-Neighbor Aggregation: PN-GNN proposes aggregating neighbor node embeddings along reasoning paths on top of conditional message passing, enhancing the logical rule expressiveness of GNNs (strictly beyond C-GNN) in a plug-and-play manner, while avoiding the generalization degradation caused by the labeling trick. The method achieves improvements on both synthetic datasets and real-world knowledge graph reasoning tasks.
Feature-Centric Unsupervised Node Representation Learning Without Homophily Assumption: This paper proposes FUEL, a method that adaptively learns the degree of graph convolution usage through a node-feature-centric clustering scheme, achieving high-quality unsupervised node representations on both homophilic and non-homophilic graphs without any homophily assumption.
Format as a Prior: Quantifying and Analyzing Bias in LLMs for Heterogeneous Data: This paper presents the first systematic investigation of format bias in LLMs when processing heterogeneous-format data (text / table / infobox / knowledge graph). Through a three-stage experimental framework, it reveals the existence of such bias, its data-level driving factors, and its internal causes at the attention mechanism level, and validates the effectiveness of attention rebalancing as an intervention.
GCL-OT: Graph Contrastive Learning with Optimal Transport for Heterophilic Text-Attributed Graphs: This paper proposes GCL-OT, the first framework to introduce Optimal Transport (OT) into graph contrastive learning for heterophilic text-attributed graphs. Three dedicated modules — RealSoftMax similarity estimation, a filter-prompt mechanism, and OT-guided latent homophily mining — address three multi-granularity heterophily challenges: partial heterophily, complete heterophily, and latent homophily, respectively.
GSAP-ERE: Fine-Grained Scholarly Entity and Relation Extraction Focused on Machine Learning: This paper introduces GSAP-ERE — a fine-grained scholarly entity and relation extraction dataset for the machine learning domain, comprising 10 entity types and 18 relation types, with 63K entities and 35K relations annotated across 100 full-text papers. Experiments show that fine-tuned models (NER: 80.6%, RE: 54.0%) substantially outperform LLM prompting approaches (NER: 44.4%, RE: 10.1%).
GT-SNT: A Linear-Time Transformer for Large-Scale Graphs via Spiking Node Tokenization: GT-SNT is proposed to leverage spiking neural networks (SNNs) as a graph node tokenizer. By aggregating multi-step propagated features into compact spike-count embeddings as node tokens, and employing Codebook-Guided Self-Attention (CGSA) to capture global context in linear time, GT-SNT achieves competitive performance on 9 node classification benchmarks while delivering up to 130× inference speedup.
Human Cognition Inspired RAG with Knowledge Graph for Complex Problem Solving: This paper proposes CogGRAG, a human cognition-inspired knowledge graph-based RAG framework that substantially improves LLM accuracy and reliability on complex Knowledge Graph Question Answering (KGQA) tasks through three stages: top-down mind map decomposition, hierarchical structured retrieval, and dual-LLM self-verification reasoning.
Hyperbolic Continuous Structural Entropy for Hierarchical Clustering: This paper proposes HypCSE, which relaxes discrete Structural Entropy (SE) into a Continuous Structural Entropy (CSE) defined in hyperbolic space. Combined with graph structure learning and contrastive learning, HypCSE enables end-to-end differentiable hierarchical clustering and consistently outperforms both discrete and continuous hierarchical clustering methods across 7 datasets.
Kernelized Edge Attention: Addressing Semantic Attention Blurring in Temporal Graph Neural Networks: This paper proposes KEAT (Kernelized Edge Attention for Temporal Graphs), which addresses the semantic attention blurring problem caused by the entanglement of node and edge representations in temporal graph neural networks. By modulating edge features with continuous-time kernels (Laplacian, RBF, and learnable MLP), KEAT achieves up to 18% MRR improvement over DyGFormer and 7% over TGN on link prediction tasks.
Logical Characterizations of GNNs with Mean Aggregation: This paper provides a systematic logical characterization of GNNs using mean aggregation: under the non-uniform setting, mean-GNNs are equivalent to Ratio Modal Logic (RML); under the uniform setting (relative to MSO), they are equivalent to Modal Logic (ML); when the combination function is additionally required to be continuous and the classification function is a threshold, the expressiveness drops significantly to Alternation-Free Modal Logic (AFML).
Magnitude-Modulated Equivariant Adapter for Parameter-Efficient Fine-Tuning of Equivariant Graph Neural Networks: This paper proposes MMEA (Magnitude-Modulated Equivariant Adapter), a lightweight parameter-efficient fine-tuning method for spherical-harmonic-based equivariant GNNs. By employing scalar gating to independently modulate feature magnitudes along "order–multiplicity" channels, MMEA achieves state-of-the-art molecular potential energy prediction accuracy—surpassing both ELoRA and full fine-tuning—while strictly preserving equivariance and using fewer trainable parameters.
MoToRec: Sparse-Regularized Multimodal Tokenization for Cold-Start Recommendation: MoToRec reformulates multimodal recommendation as a discrete semantic tokenization task. By leveraging a sparsely-regularized Residual Quantization VAE (RQ-VAE), raw multimodal features are transformed into composable discrete semantic codes. Combined with adaptive rarity amplification and a hierarchical multi-source graph encoder, the framework effectively addresses the item cold-start problem.
MUG: Meta-path-aware Universal Heterogeneous Graph Pre-Training: MUG is the first universal heterogeneous graph pre-training method that requires no LLM. It unifies heterogeneous node/relation types via contextual structural encoding, aligns representation spaces across graphs with a dimension-aware encoder, and achieves transferable encoding and aggregation through a shared GNN encoder over meta-path views combined with global scatter regularization. MUG substantially outperforms existing methods on cross-domain and few-shot node classification.
MyGram: Modality-aware Graph Transformer with Global Distribution for Multi-modal Entity Alignment: This paper proposes MyGram, which captures deep structural contextual information within each modality via a Modality-aware Graph Diffusion (MGD) module, and introduces a global distribution alignment loss (Gram Loss) based on the determinant of the Gram matrix to enforce cross-modal semantic consistency in high-dimensional space, achieving more robust multi-modal entity alignment.
NOTAM-Evolve: A Knowledge-Guided Self-Evolving Optimization Framework with LLMs for NOTAM Interpretation: This paper proposes NOTAM-Evolve, a self-evolving framework that achieves dynamic knowledge grounding via knowledge graph-enhanced tabular retrieval (KG-TableRAG), combined with iterative SFT+DPO preference optimization and a multi-view voting inference mechanism. The framework enables a 7B-parameter LLM to autonomously master deep parsing of complex aviation NOTAMs, achieving a 30.4% accuracy improvement over the base LLM.
NTSFormer: A Self-Teaching Graph Transformer for Multimodal Isolated Cold-Start Node Classification: This paper proposes NTSFormer (Neighbor-to-Self Graph Transformer), a unified Graph Transformer framework that implements a self-teaching paradigm via a cold-start attention mask. Within the same model, a "student" prediction is derived solely from the node's own features while a "teacher" prediction leverages neighbor information, enabling end-to-end self-teaching without degrading to an MLP. The framework handles missing modalities in multimodal graphs through MoE input projection and multimodal graph pre-computation.
On Stealing Graph Neural Network Models: This paper demonstrates that under strict query budgets (e.g., only 100 queries), an attacker can efficiently steal a GNN model via a two-stage approach: (1) locally obtaining an encoder (randomly initialized or SSL-trained) without interacting with the victim, and (2) strategically selecting queries via K-means clustering. On the Physics dataset, the proposed method achieves 91% accuracy with only 100 queries, whereas the current state-of-the-art requires approximately 5,000 queries plus additional access to victim embeddings to reach comparable performance.
PathMind: A Retrieve-Prioritize-Reason Framework for Knowledge Graph Reasoning with Large Language Models: This paper proposes PathMind, a framework following the Retrieve-Prioritize-Reason paradigm. It identifies important reasoning paths via a semantics-aware path prioritization function that jointly considers cumulative cost and estimated future cost (inspired by A*), and then enhances faithful and interpretable LLM reasoning through a two-stage training strategy comprising task-specific instruction tuning and path-level preference alignment. PathMind achieves state-of-the-art performance on complex reasoning tasks while consuming significantly fewer tokens.
PCoKG: Personality-aware Commonsense Reasoning with Debate: This work constructs PCoKG, the first large-scale personality-aware commonsense knowledge graph comprising 521,316 quadruples \((e, p, r, t)\) (event–personality–reasoning dimension–tail), generated via LLM role-playing combined with a multi-agent debate mechanism to produce high-quality personality-differentiated inferences. Experiments validate that MBTI personality information enhances commonsense reasoning and personalized dialogue generation.
Posterior Label Smoothing for Node Classification: This paper proposes PosteL (Posterior Label Smoothing), which derives soft labels from neighborhood label distributions via Bayesian posterior inference for node classification. The method naturally adapts to both homophilic and heterophilic graphs, achieving accuracy improvements in 76 out of 80 combinations across 8 backbone architectures and 10 datasets.
Relink: Constructing Query-Driven Evidence Graph On-the-Fly for GraphRAG: This paper proposes a paradigm shift in GraphRAG from "build-then-reason" to "reason-and-construct," introducing the Relink framework that dynamically constructs query-specific evidence graphs—combining a high-precision KG backbone with a high-recall latent relation pool, unified via a query-driven ranker to assess relevance on demand, complete missing paths, and filter distractor facts—achieving average gains of 5.4% EM and 5.2% F1 across 5 multi-hop QA benchmarks.
RFKG-CoT: Relation-Driven Adaptive Hop-count Selection and Few-Shot Path Guidance for Knowledge-Aware QA: This paper proposes RFKG-CoT, which enhances LLM reasoning over knowledge graphs via two components: relation-driven adaptive hop-count selection (dynamically adjusting reasoning steps using KG relation activation masks) and few-shot path guidance (in-context examples in a Question-Paths-Answer format). Evaluated on 4 KGQA benchmarks, the method achieves significant improvements — GPT-4 reaches 91.5% (+6.6pp) on WebQSP, and Llama2-7B gains up to +14.7pp.
S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning: This paper proposes S-DAG, which uses a GNN to identify relevant subjects and their dependencies from a given question, constructing a directed acyclic graph. Subject nodes are matched to the most capable expert LLMs (14 domain-specific models of 7–13B parameters), and collaborative reasoning proceeds in DAG topological order (supporting subjects → dominant subject). The resulting small-model pool surpasses GPT-4o-mini (59.73 vs. 58.52) and approaches the performance of a 72B model.
Self-Adaptive Graph Mixture of Models: This paper proposes SAGMM (Self-Adaptive Graph Mixture of Models), a graph MoE framework that leverages architectural diversity by employing a Topology-Aware Attention Gating (TAAG) mechanism to adaptively select and combine heterogeneous GNN experts, coupled with an adaptive pruning mechanism. SAGMM consistently outperforms individual GNNs and existing MoE methods across 16 benchmarks spanning node classification, graph classification, regression, and link prediction.
Self-Correction Distillation for Structured Data Question Answering: This paper proposes Self-Correction Distillation (SCD), which transfers structured data question answering capabilities from large-scale LLMs (GPT-4) to small-scale LLMs (8B) via an Error Prompting Mechanism (EPM) and a two-stage distillation strategy, achieving state-of-the-art distillation performance across five benchmarks.
Sentient: Detecting APTs Via Capturing Indirect Dependencies and Behavioral Logic: This paper proposes Sentient, an APT detection method combining Graph Transformer pre-training and bidirectional Mamba2 intent analysis. Trained exclusively on benign data, it captures indirect dependencies, removes contextual noise, and correlates behavioral logic, achieving an average 44% reduction in false positive rate across three standard benchmarks.
Sheaf Graph Neural Networks via PAC-Bayes Spectral Optimization: This paper proposes SGPC (Sheaf GNNs with PAC-Bayes Calibration), which integrates Wasserstein optimal transport for learning sheaf restriction maps, variance-reduced diffusion with an adaptive frequency mixing layer, and PAC-Bayes spectral regularization. SGPC consistently outperforms existing GNN and sheaf methods on both homophilic and heterophilic graph node classification benchmarks while providing theoretical generalization guarantees.
Spiking Heterogeneous Graph Attention Networks: This paper proposes SpikingHAN, the first framework to introduce Spiking Neural Networks (SNNs) into heterogeneous graph learning. It employs a single-layer graph convolution with shared parameters to aggregate meta-path-based neighborhood information, fuses multiple meta-path semantics via semantic-level attention, and encodes the resulting representations into 1-bit binary spike sequences. SpikingHAN achieves competitive node classification performance on three datasets with fewer parameters, faster inference, and lower energy consumption.
UniHR: Hierarchical Representation Learning for Unified Knowledge Graph Link Prediction: This paper proposes UniHR, a unified framework that converts hyper-relational, temporal, and nested KGs into a triple-based representation via Hierarchical Data Representation (HiDR), and designs a Hierarchical Structure Learning (HiSL) module for two-stage intra-fact and inter-fact message passing. UniHR achieves state-of-the-art or competitive link prediction results across 9 datasets spanning 5 KG types.