Showing 45 papers for 2026-05-26
This work introduces a fact generation task for hyper-relational knowledge graphs and proposes a generative representation learning method based on masked discrete diffusion to fill multiple missing components of a fact. It extends beyond standard link prediction by handling when several or all constituents of a fact are unknown.
Biological systems are governed by structured molecular interactions; the authors propose using graph-encoded domain knowledge to improve learning when data are scarce. They present Graph-in-Graph (GiG), a framework that modulates deep learning with the external knowledge graph to preserve relevant structure and boost predictive performance in limited-sample clinical studies.
As large foundation models provide strong priors, naive aggregation in multimodal graphs can introduce topological noise and hurt performance. The paper proposes Prior-Retaining Decoupled Learning to separate LFM priors from graph aggregation in MAGL, enabling robust learning without overfitting to topology.
MedMamba integrates multi-scale convolutional embeddings with state-space models and adaptive graph learning to model local-global dynamics and nonstationarities in medical time series. It is end-to-end and aims to capture latent channel interactions for classification.
TGFormer is a Transformer designed for temporal graphs using an auto-correlation mechanism. It reframes temporal graph learning as a trajectory problem aligned with time-series, addressing long-term dependencies and periodic patterns.
GL-LFGNN addresses EEG emotion recognition with a global-local dual-branch causal GNN that models directed information flow using Liang-Kleeman causal measures rather than symmetric adjacency. The model aims to better reflect asymmetric neural information transfer.
Revisiting Pre-Propagation GNNs studies diffusion-then-transformation decoupled architectures. The authors propose robust diffusion operators and hidden-state re-propagation to improve expressivity, bridging the gap to message-passing GNNs.
This work studies diffusion-transformer frameworks for unsupervised constraint-optimization. While naive implementations struggle with general discrete variables and global constraints, the authors propose a Blocked Gibbs approach to enable scalable, global reasoning with diffusion transformers.
ASTRO introduces adaptive spatio-temporal reinforcement optimization for GNN-powered anomaly detection in IIoT CPS. It models spatio-temporal relationships and uses reinforcement learning to optimize detection performance against cyber-physical threats.
3C framework captures learners' perceived knowledge from open-ended self-reports, builds a heterogeneous graph linking learners and knowledge, calibrates estimates, and coaches with adaptive feedback to support self-regulated learning.
Si'multaneous STMP proposes a joint spatial-temporal message passing method for dynamic graphs, avoiding the traditional sequential separation of spatial and temporal steps. This enables simultaneous reasoning over structure and its evolution, improving dynamic representation.
We investigate whether deterministic closed-form solvers can match the performance of trained GCN-like models on node classification. The authors introduce a routed closed-form framework guided by adjusted graph homophily, yielding exact unlearning capabilities and competitive results on assortative graphs.
The authors model professional football passing decisions as a dynamic graph problem and develop an enhanced MPNN approach to Receiver Selection. Players are nodes with spatial/context features, passes are edges with distance and pressure, to predict optimal target.
This work proposes invariant-based weight sharing for MPNNs by indexing weights with graph invariants—permutation-preserving features—allowing weights to be shared across structurally equivalent subgraphs.
A comprehensive book surveying fuzzy, neutrosophic, and related uncertainty models in graph theory, covering concepts, properties, classes, and parameters, and discussing practical applications such as uncertain molecules.
BoxLitE embeds knowledge base concepts into convex regions to faithfully capture hierarchies in TBoxes, with region sizes reflecting generality. It provides a convex-optimization-based approach to KB embeddings.
The paper argues for integrating deep learning design directly into database query systems, avoiding round-trips to external ML frameworks. It discusses building DL on relational data via in-database architectures that reason over joins and tuple embeddings.
Introduce Gaussian rank-based neighborhood degree normalization for GNNs in image classification. This approach weights neighbor contributions by a Gaussian-rank of their degrees to improve semi-supervised learning where graphs reflect image structure.
Deep Graph Laplacian Regularization (Deep GLR) adds graph-based quadratic regularization to LDCT reconstruction under strict parameter limits, enabling accurate reconstructions with far fewer parameters.
The article reviews multiple statistical perspectives on GNN generalization, distinguishing three broad frameworks: learning-theory-based uniform convergence, hypothesis complexity, and alternative perspectives to understand performance on graphs.
We provide theoretical and empirical evidence that message-passing GNNs cannot approximate sparse triangular factorizations for matrix classes that need nonlocal dependencies to yield high-quality preconditioners. We build baselines using synthetic matrices and real-world SuiteSparse examples to illustrate the limitation.
HiTeC proposes Hierarchical Contrastive Learning on Text-Attributed Hypergraphs with Semantic-Aware Augmentation to capture correlations between textual semantics and hypergraph topology. It addresses limitations of graph-agnostic text encoders by integrating semantics with topology at multiple levels.
HiGraph introduces the largest public hierarchical graph dataset for malware analysis, with over 200M CFGs that expose hierarchical structure linking high-level functional interactions to low-level instruction logic. The dataset enables researchers to study malware programs as multi-scale graphs.
We present CF-HyperGNNExplainer, a counterfactual explanation method for HGNNs that finds minimal structural edits to flip a prediction. It constructs counterfactual hypergraphs by actionable edits such as removing node–hyperedge incidences or deleting hyperedges, yielding concise, actionable explanations.
ORACAL is a robust and explainable multimodal framework for smart contract vulnerability detection that uses causal graph enrichment to capture the interplay of control flow and data dependencies. This integration improves robustness to adversarial attacks and provides explainable evidence.
Full-Spectrum GNNs (FSpecGNNs) extend spectral GNNs to second-order signals by lifting features to the node-pair domain and expanding spectral filtering. They achieve higher expressivity beyond the 1-WL bound while remaining scalable.
HypergraphFormer learns hypergraph representations from LLMs to generate editable floor plans. Trained by supervised fine-tuning to produce a hypergraph textual representation encoding spatial relations and connectivity on the RPLAN dataset, with demonstrated generalization to out-of-distribution data.
We analyze Deep Neural Sheaf Diffusion (NSD) and show that while NSD provides theoretical guarantees against collapse, in deep settings the disagreement signal of the sheaf Laplacian decays, leading to diminishing contribution from deeper layers. The paper identifies practical limits and suggests remedies.
Graph Navier Stokes Networks (GNSN) draw inspiration from Navier–Stokes equations to overcome oversmoothing in diffusion-based GNNs by incorporating convection into message passing. This enables deeper and more expressive models.
The paper develops a spectral framework for neural operators on graphs using graphons as limits, establishing convergence guarantees and transferability under various regularity assumptions (no regularity, global Lipschitz, piecewise-Lipschitz), and linking graph convergence to operator convergence.
Universal Graph Backdoor Defense presents a feature-based homophily perspective showing that current structure-centric defenses fail against emerging backdoor attacks; proposes a unified defense leveraging node features and local homophily to detect and mitigate GBAs.
Advancing Graph Few-Shot Learning via In-Context Learning proposes using in-context learning to perform graph few-shot classification without heavy fine-tuning and without task adaptation during inference, leveraging unlabeled nodes to boost performance. The approach emphasizes efficiency and scalability for practical graph applications.
Clustering as Reasoning reframes chain-of-thought graph learning as a k-means process, offering a k-means interpretation of iterative reasoning on TAGs; critiques disjoint architectures and provides more integrated semantic-topological interaction.
Neural Scalable Symbolic Search Framework for Complex Logical Queries with Multiple Free Variables tackles existential first-order queries with k free variables on incomplete KGs, and provides scalable neural-symbolic search for multi-variable queries, moving beyond marginal rankings.
L2IR reveals latent intent in graph fraud detection; fraudsters create many connections to benign users causing signals to be diluted; while LLMs provide semantic cues, L2IR aims to reveal latent intent behind suspicious connections.
Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience explores whether KG-driven reasoning can emerge in neuroscience using a high-quality KG distilled from a single authoritative textbook, enabling KG-grounded QA supervision to support expert-level reasoning.
Adaptive Graph Refinement and Label Propagation with LLMs for Cost-Effective Entity Resolution proposes using LLMs to adaptively refine graphs for dirty ER, addressing blocking failures and mislinks by creating better edges and propagation for improved clustering.
Efficient and Scalable Neural Symbolic Search for Knowledge Graph Complex Query Answering discusses efficiency bottlenecks of neural-symbolic CQA and proposes an approach that reduces quadratic data complexity and NP-hard query complexity, enabling scalable CQA on large KGs.
Leveraging Spreading Activation for Improved Document Retrieval in Knowledge-Graph-Based RAG Systems notes that RAG systems treat retrieved information as equally reliable; proposes using spreading activation over a knowledge graph to model credibility and interconnectedness, improving document retrieval.
Knowledge Graph Re-engineering Along the Ontological Continuum (extended version) discusses diversity of KG modelling from lightweight vocabularies to rich ontologies, arguing for principled KG reengineering to fit neuro-symbolic AI needs; GenAI offers automation but requires principled understanding.
This paper presents a stacking ensemble of attention-guided Graph Neural Networks for malware detection on control-flow graphs. It emphasizes interpretability and robustness to evasion by providing attention-based explanations and combining multiple models to improve generalization across diverse malware families. By integrating an ensemble with explainable signals, the approach aims to outperform single-model GNNs in security tasks.
SentGraph introduces a sentence-level graph based retrieval augmented generation framework for multi-hop question answering. It builds a hierarchical sentence graph across retrieved documents to form coherent evidence chains, addressing the noise and incoherence of chunk-based retrieval. The approach improves evidence organization and enables more reliable multi-hop reasoning and answer generation.
The paper presents a repeatable structured expert elicitation methodology and a federated Semantic Knowledge Graph (SKG) architecture to capture tacit lab workflow knowledge, including failure conditions and decision logic. It enables querying and cross-site sharing of expert knowledge, improving reuse and governance of laboratory workflows. The approach is demonstrated in a Biochemical and Cellular Pharmacology department, showcasing practical deployment in pharmaceutical research.
PiPNN, or Pick-in-Partitions Nearest Neighbors, is an ultra-scale graph-based indexing method for approximate nearest neighbor search. It significantly reduces construction time compared with state-of-the-art indexes like HNSW and Vamana by introducing HashPrune, an online pruning technique that cuts computational cost during graph construction while preserving high query performance. This enables ultra-large-scale NN search suitable for industrial datasets.
MetaboKG is an analysis-centric knowledge graph framework that integrates untargeted metabolomics data, including MS/MS spectra, features, workflow outputs, annotations, confidence evidence, and contextual metadata. It addresses fragmentation in the analytical layer by enabling unified querying, reuse, and reproducibility across repositories. Building on ecosystems such as Pan-ReDU, ENPKG, and METRIN-KG, MetaboKG connects datasets to support cross-study metabolomics analysis.