Showing 48 papers for 2026-06-23
We propose CIExplainer++, a perturbation-based method grounded in causal inference to explain Graph Neural Networks. It identifies subgraphs that have genuine causal influence on the model’s outputs and translates these structures into interpretable representations.
We introduce MMGNN, a multi-level, multi-color GNN framework for molecular property prediction. It decomposes a molecular graph into overlapping atom-pair specific subgraphs while preserving atom-level detail; MMGNN-2D uses color subgraphs from covalent connectivity and MMGNN-3D incorporates geometric information.
We propose DS-HGNN, a Dual-Stream Heterogeneous Graph Neural Network for fast and accurate full-field predictions of stress and displacement in stiffened panels. The model handles varying topology and boundary conditions and serves as a fast surrogate to finite element analysis.
We introduce hierarchical pooling for Sheaf Neural Networks to enable multi-resolution analysis. Sheaf neural networks generalize GNNs by using stalk-valued signals and restriction maps; existing architectures operate at fixed resolutions, and hierarchy enables scalable modeling across scales.
We present a Gated Graph Neural Network (GGNN) approach for fast-convergent dynamic average estimation in distributed multi-agent systems. The method leverages the GGNN structure to achieve robust convergence despite topology changes, enabling fully decentralized operation. Experiments show faster convergence and robustness.
We propose Federated Temporal Graph Convolutional Network with Advantage Actor-Critic (Federated TGCN-A2C) for privacy-preserving cyber-resilient IoMT. The architecture integrates PyG-based TGCN with an anomaly gate and four defense mechanisms, including digital-twin guided adaptive deception to counter threats.
We show memory-efficient graph filtering can scale collaborative filtering by smoothing graph signals with polynomial filters. Training-free GF-based CF reduces memory and computation while achieving competitive accuracy.
We propose a framework for learning Directed Acyclic Hypergraphs (DAHGs) from observational data, enabling joint parental influences beyond pairwise relationships. The approach uses a generalized linear SEM with multiplicative interaction terms, where non-zero weights indicate directed hyperedges.
We present DevoTG, temporal graph neural networks for C. elegans development, using two representations: a Continuous-Time Dynamic Graph of cell divisions and a Discrete-Time Dynamic Graph of evolving synaptic connectome. The framework predicts developmental trajectories and synaptic changes.
We propose a generic, model-agnostic pipeline for local counterfactual explanations in GNNs. It couples factual explainability with missing-edge prediction models from link prediction to improve quality, robustness, and consistency of counterfactuals in graphs.
We study speed-quality tradeoffs of GNNs for link prediction by introducing early-exit strategies without extra auxiliary losses. Exiting decisions emerge implicitly during inference, enabling faster predictions with controllable accuracy.
We introduce multigrid training for graph-based molecular generation to accelerate learning across resolutions. By transferring parameters across low- and high-resolution discretizations, training at coarse scales speeds up and stabilizes optimization for high-resolution graphs.
We propose weight-level information injection to enhance LLMs for graph tasks by graph-aware LoRA generation, improving how LMs encode whole-graph information and enabling better graph understanding.
We provide theoretical analysis showing step dependence is necessary and sufficient for adaptive recurrent message passing at test time. Recurrent graphs enable adapting computation to data distribution shifts, improving efficiency.
GRADE models any LLM agent run as a graph of steps with two edge layers: execution edges and dependency edges. It can log what ran and what it relied on, inferring missing dependencies and grading the edges by their evidential support across six corpora.
We propose Field-informed Graph Network (FGN) where latent information entropy fields govern edge formation and message passing, enabling graphs to emerge from a latent explanatory field.
We introduce PromptDyG, test-time prompt adaptation for dynamic graphs to handle evolving snapshot complexities. Unlike offline learning, prompts are adapted at test-time to capture changing graph dynamics.
We propose TaLK, distillation for text-attributed graphs by coupling a language model with a graph-aware kernel to distill informative datasets, reducing training costs for TAGs.
We propose quaternion representation learning for clustering heterogeneous attributed graphs, addressing attribute heterogeneity and representation degradation. The quaternion-based approach fuses attributes and topology to produce robust clusters.
We formalize chem-aware admission as a pipeline gate (ChemGuard) to test whether submitted molecular records survive parsing, canonicalization, and consistency checks in real-world learning pipelines, evaluating backdoor threats in a realistic setting. This complements existing defenses.
This work develops a spectral theory for normalized corrected GNN propagation by analyzing the symmetric normalized adjacency with its degree-stationary component removed, aligning with the normalization used in standard GCN-like models while isolating the stationary direction linked to oversmoothing. The main result is a high-probability exact-recovery theorem for binary contexts, showing the corrected operator preserves discriminative signal across many propagation layers.
FeLoG presents a scalable distributed graph embedding framework that uses a feedback loop mechanism to align sampling with embedding quality in large graphs. It reduces redundant exploration and improves embedding quality and efficiency at scale.
ThermoLLM introduces a thermodynamics-aware HVAC control framework that leverages a spatial-semantic knowledge graph to reason about zone coupling, adjacency, and delayed thermal interactions. This enables LLM-based controllers to make thermodynamically informed decisions beyond description-only prompts.
SuperCond-GNN proposes a graph neural network surrogate to predict voltage distributions in high-temperature superconducting magnets. By mapping lumped-element circuits to graphs, the model learns electrical responses as functions of topology, material properties, and operating current, demonstrated on tape stacks up to ten.
Provable Filter for Real-world Graph Clustering addresses the limitations of clustering on heterophilic graphs and proposes a principled, theoretically grounded filtering approach that works well on real-world data. The method offers guarantees under practical conditions.
A polarity-aware multi-relational model (PAMR) is introduced to predict signed interactions in biological networks, distinguishing polar interactions such as activation and inhibition. The model handles multiple relation types with polarity awareness to improve pharmacological predictions.
This work reevaluates heterophilic graph learning and highlights pitfalls in heterophily benchmarks and homophily metrics. It calls for more robust evaluation and critical assessment of heterophilic GNNs beyond existing benchmarks.
Mixture-of-Experts Graph Transformers (MoE-GT) are proposed to improve interpretability in particle collision detection by deploying multiple experts over graph-structured collider data, enabling more transparent decision processes.
Structure-Aware Compound-Protein Affinity Prediction uses Graph Neural Networks with Group Lasso regularization to produce explainable predictions and identify key structural motifs driving binding.
FedSA-GCL introduces a semi-asynchronous federated graph learning framework with personalized aggregation and cluster-aware broadcasting to improve efficiency and convergence in heterogeneous subgraphs.
Turning Tabular Foundation Models into Graph Foundation Models proposes methods to adapt tabular foundation models for graphs by handling arbitrary node features and enabling in-context learning on graph data.
Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication presents a method that uses dual-pass spectral encoding to preserve node distinctiveness and a latent-space communication step to align latent spaces across graphs, improving unsupervised alignment.
GraphPFN introduces a Graph Foundation Model built on prior-data fitted networks, pretrained on synthetic graphs to enable robust predictions in data-scarce regimes with in-context learning.
VDW-GNNs define vector diffusion wavelets and integrate them into geometric GNNs, enabling geometry-aware representations and showing effectiveness on synthetic and real-world geometric data, with supporting theoretical properties.
CQD-SHAP provides explainable complex query answering by applying Shapley values to CQD reasoning steps, enabling attribution of contributions to multi-hop queries over incomplete knowledge graphs.
Temporal Graph Pattern Machine (TGPM) learns transferable temporal evolution patterns in dynamic graphs, addressing limitations of short-term dependency modeling and static neighborhood semantics for more general temporal reasoning.
GraphSVR couples graph convolution with support vector regression to robustly forecast urban air pollution, capturing spatiotemporal dependencies while handling nonlinearities and anomalous observations.
O-RAN Xapps Conflict Prediction uses graph convolutional networks to forecast conflicts among xApps in O-RAN deployments, enabling proactive coordination and performance protection.
A Hybrid TGN-SEAL model combines Temporal Graph Networks with SEAL-style local subgraph reasoning to improve dynamic link prediction in evolving networks, addressing sparsity and class imbalance.
WarPGNN presents a parametric thermal warpage analysis framework powered by graph neural networks for chiplet-package systems, offering fast, scalable physics-informed warpage predictions for design-space exploration.
We introduce dualGNN, an autoregressive message-passing GNN for sampling fine, regular triangulations of lattice polytopes. It operates on a generalized dual graph whose edges are labeled by signed circuits from oriented matroids, and we show these circuits are necessary and sufficient to determine a triangulation's regularity when magnitude information is retained. The approach is independent of the polytope's point count and robust to representation choices.
We analyze the use of LLMs augmented with knowledge graphs in materials discovery and identify contextual tunneling, where models over-anchor narrow retrieved evidence at the expense of global mechanistic reasoning. To fix this, ARIA is a causal-aware framework that conditions knowledge use on mechanistic completeness, guiding retrieval and reasoning to preserve causal consistency.
We propose a production-grade GraphRAG stack that enforces verifiable graph chain-of-thought paths for low-hallucination multi-hop medical reasoning on a ~700K-node medical knowledge graph powering a fertility assistant. A distance-indexing PLL oracle provides exact distances for fast feasibility checks, while a lightweight AStarNet heuristic stays inside the PLL corridor to enumerate simple paths.
We present a human-in-the-loop framework for Automated Semantic Fault Localization in SysML v2 that detects semantic errors beyond compiler checks. It combines a fine-tuned small language model with a domain knowledge graph encoding physical compatibility to locate faults and propose repairs, with humans in the loop to validate.
We introduce ARLtR, a unified framework for automated knowledge graph creation and question generation that bridges vector-based retrieval and reasoning over knowledge graphs. It yields ground-truth entities, relations, and fact-grounded questions aligned with the underlying corpus, enabling end-to-end QA and KG construction.
We propose SHACR, a graph-augmented semi-autonomous framework for resolving multi-class conflicts in smart home IoT automation. By modeling cross-rule interactions among heterogeneous devices and environmental factors as a graph, SHACR detects unsafe, wasteful, or privacy-threatening behaviors that are invisible to text-only analysis and can suggest or enact safe resolutions.
We address the lack of compositionality in graph query languages like GQL and SQL/PGQ by proposing a compositional language for property graphs, together with theoretical foundations and a route to standardization. We show current languages handle reachability and first-order queries but struggle with more complex compositions and higher complexity classes such as NLOGSPACE.
We conduct a controlled study on clinical QA to test whether grounding LLMs with knowledge graphs helps. We reproduce findings that KG grounding yields benefits mainly for out-of-training knowledge, while in-domain knowledge does not consistently improve performance and can even hurt, highlighting when grounding is advantageous.