Showing 42 papers for 2026-07-07
Proposes PGRE (Poisson-Gamma Relational Evolution), a probabilistic model for dynamic knowledge graphs that captures temporal and inter-relational dependencies. It represents multi-relational dynamics under noise and incompleteness to improve downstream tasks.
Introduces MolBasic, a structure-first framework that translates canonical SMILES into molecular graphs to ground structure in LLM reasoning. This addresses the gap where molecular LLMs struggle to recognize basic molecular structures.
Offers a heterogeneous graph condensation method based on role-aware clustering to summarize large heterogeneous graphs before HGNN training. It avoids gradient matching and bilevel optimization, providing a practical path for scalable HGNN learning.
Presents PhenoNEST, a neuro-symbolic framework for building an ontology-aware multimodal knowledge graph from noisy field notes and RGB images to study genotype-phenotype interactions over time. The pipeline extracts entities and relations and drives multimodal reasoning.
Introduces a near-linear-time solver for graph p-Laplacian semi-supervised learning by continuing in p. The approach turns the nonlinear p-Laplacian problem into a sequence of tractable solves for p>d, enabling efficient SSL with limited labels.
Proposes NetinfoGC, a graph classification framework that extends Network Usable Information (NUI) to graphs. It creates a family of permutation-invariant representations from propagation-based signals and classic structural descriptors, and includes a training-free NUI estimation method based on clustering to guide representation quality and structure selection.
Foundations of Equivariant Deep Learning introduces order-equivariant neural networks (OENN) that unify graph and sheaf neural networks using equivariant bundles over face posets. It characterizes all linear order-equivariant maps and builds practical OENN layers.
Develops a target-aware, interaction-guided reinforcement learning attack for black-box node injection on GNNs, jointly learning malicious node features and edges to improve attack efficacy under budget constraints.
Introduces PGU-OD, a physics-informed graph learning framework that propagates uncertainty across feature extraction, topology learning, and decision making to tackle open-set domain generalization in fault diagnosis.
Proposes a graph-transformer approach on superpixel graphs for image classification that preserves geometrical invariances, reducing redundancy and improving robustness to geometric transformations.
PREDIKTOR is a patient-centered multi-view framework that aligns a pre-treatment patient-specific knowledge graph with gene-level perturbation representations from drug-induced expression changes to predict therapeutic outcomes.
Offers a unified quantitative benchmarking framework for GNN explainability (G-XAI) requiring no ground-truth and formalizing graph explainability metrics for practical evaluation and method selection.
Explores personalized differential privacy for decentralized local graphs, proposing tailored LDP mechanisms that enable private graph data collection and learning without compromising individual node privacy.
Examines hyperparameter transfer in graph neural networks, showing that near-optimal hyperparameters can transfer across model scales, and proposes strategies to proxy-tune on small models to guide large-model training.
FAST is a holistic framework that optimizes memory I/O, computation, and temporal neighbor sampling to speed up end-to-end Temporal GNN training on large dynamic graphs.
GeoFlow augments area representations with geospatial attributes and uses a geometric-intrinsic fusion encoder to jointly model intrinsic area signals and spatial relations for improved origin-destination flow prediction and generation.
Presents scalable GPU-based algorithms to compute 1-WL expressiveness on massive graphs, enabling practical WL expressiveness analysis on real-world graphs with parallel processing.
Introduces a GNN-based method to synthesize schedulable Job-Level Dependencies for cause-effect chains in mixed-criticality automotive software, enabling data-age bounding independent of scheduling decisions.
Studies Graph Neural Networks for the graphical bootstrap approach on over 20 million graphs from planar N=4 SYM calculations, showing robust generalization to larger graphs and potential speedups over traditional bootstrap algorithms.
KARMA formalizes a knowledge-graph-based automated reasoning materialization and alignment framework that enumerates schema-constrained paths into slot-aligned contrastive candidates; Slot-Parallel Alignment (SPA) then applies a slot-level objective to route supervision to discriminative entity slots.
GlaKG introduces a biomarker-centric fundus knowledge graph for glaucoma diagnosis and risk assessment. It encodes six entity types (Fundus Image, Optic Disc, Neural Rim, Pathology, Diagnosis, Risk Level), eight relation types, and 11 clinically grounded rules to integrate biomarkers, clinical knowledge, and image features for traceable reasoning. The framework enables more transparent glaucoma diagnosis and risk stratification.
Graph representation learning of longitudinal medical imaging trajectories for treatment response prediction. It proposes an imaging-based 3D spatio-temporal framework that models longitudinal imaging trajectories to estimate individualized responses to neoadjuvant chemotherapy in breast cancer, with a focus on predicting pathological complete response (pCR). The method aims to support personalized treatment decisions.
TACTIC-KG Toward Small Agent Teams for Cyber Threat Intelligence Knowledge Graph Construction. Cyber Threat Intelligence reports are unstructured and noisy, and large LLMs are costly and unstable. The paper proposes a multi-agent approach using small agents to collaboratively extract and assemble CSKGs, offering better controllability and cost efficiency.
Graph Unitary Message Passing GUMP introduces a unitary propagation operator on transformed graphs to avoid graph-induced exponential decay during repeated propagation. By mapping the input graph to an Eulerian line-graph, it enables stable deep message passing.
FLASH Flexible Learning of Adaptive Sampling from History in Temporal Graph Neural Networks. It replaces static historical neighbor sampling with a learnable graph-adaptive mechanism that generalizes existing heuristics. This enables TGNNs to adapt sampling to graph structure for better efficiency and accuracy.
Self-Improving Neural Pruning A Graph Neural Network Framework for Scalable Mixed Bundle Pricing. Mixed bundle pricing is combinatorially hard. We propose a GNN-guided pruning framework to identify a smaller subset of bundles for pricing, enabling scalable non-additive bundle pricing.
Graph Representation Learning Augmented Model Manipulation on Federated Fine-Tuning of LLMs. Federated fine-tuning preserves privacy but remains vulnerable to model manipulation. The paper proposes a graph representation learning augmentation to detect and mitigate adversarial updates during aggregation, improving robustness.
Selecting Samples on Graphs A Unified Dataset Pruning Framework for Lossless Training Acceleration. The framework unifies intrinsic and extrinsic pruning signals into a single method that robustly prunes across different ratios and data distributions while preserving training performance.
SpatialThinker SpatialThinker Reinforcing Scene Graph-Grounded Spatial Reasoning via Dense Rewards. Addresses spatial reasoning challenges in multimodal LLMs by unifying Scene Graph Generation and visual reasoning in a single pass through online reinforcement learning with dense rewards.
Graph Neural Networks are Heuristics. For the Euclidean Traveling Salesman Problem, a non-autoregressive GNN is trained with a differentiable Hamiltonian-cycle objective and no labels, rewards, or search. It outputs a complete tour in a single forward pass.
An Additive MLP-GNN Framework for Characterizing Chemical and Structural Contributions to Aqueous Solubility. The model keeps physicochemical descriptors and molecular graph information separate through an additive architecture, with a chemical branch (MLP) and a graph branch (GNN) to yield interpretable contributions to solubility.
Enhanced Feature Extraction for IoT Network Intrusion Detection Using GNNs and KAN. Proposes SKGFusionKAN, an IoT-tailored approach that addresses dynamic topologies, imbalanced traffic, and sparse connections by enhancing feature extraction for IoT intrusion detection with graph neural networks and knowledge-aware networks.
Conditional Diffusion Guided Knowledge Transfer for Multi-Domain Knowledge Graph Completion. Transfers knowledge from support KGs to a target KG via a conditional diffusion guided generative process while preserving domain-specific contextual information, especially in low-resource scenarios.
LLM-Enhanced Hierarchical Heterogeneous Graph Representation Learning for Malicious Python Package Detection. Combines hierarchical heterogeneous graph representations of program entities with LLMs to enhance detection of malicious PyPI packages through improved code understanding and semantic reasoning.
TRIAGE Trustworthy Retrieval Instrumentation And Graph Evaluation. Proposes stage-aware instrumentation for automated graph-RAG covering extraction, graph construction, and inference to localize failures, enabling trustworthy retrieval augmented generation.
TokenMizer Graph-Structured Session Memory for Long-Horizon LLM Context Management. Maintains session history as a typed knowledge graph and, at context boundaries, replaces the raw transcript with a token-budgeted serialization of session state including decisions and rationale.
GenShin Guiding Rational Liposome Design by Ranking Liposomal Protein Corona through a Docking-Pose-Free GNN. Proposes a docking pose free GNN to rank protein corona composition on liposomes, enabling pre synthesis screening of large lipid spaces and accelerating rational LNP design.
ARISE A Repository-level Graph Representation and Toolset for Agentic Program Repair and Fault Localization. Proposes a multi-granularity repository graph and a toolset that enables repository-scale automated program repair and fault localization in a framework-agnostic manner.
CVE-TTP KG Knowledge Graph Linking Software Vulnerabilities to Attack Behaviors. Bridges vulnerability databases with attacker behaviors by linking CVEs to tactics and techniques from MITRE ATT&CK through a knowledge graph, enabling better threat interpretation and response.
Patient-Conditioned Dual Hypergraph Reasoning for Auditable Traditional Chinese Medicine Prescription Support. Proposes patient-conditioned dual hypergraph reasoning to connect clinical narratives, syndromes, herbs and doses, enabling auditable and evidence-based Traditional Chinese Medicine prescriptions.
Hyper-KGGen is a skill-driven framework for generating high-quality knowledge hypergraphs that encode n-ary facts, aiming to advance semantic representations beyond traditional binary graphs. It addresses the scenario gap where generic extractors fail to generalize across domains with domain-specific jargon and where existing methods struggle to balance structural skeletons with fine-grained details. By reframing extraction as a set of transferable skills, Hyper-KGGen seeks to produce more accurate, domain-aware hypergraphs.
GORIO introduces GPU-Centered Remote I/O for Graph ANNS over NVMe-oF. It tackles memory bottlenecks by shifting I/O orchestration from CPU-centric paths to GPU-driven data movement over disaggregated storage, enabling efficient access to large vector indexes that do not fit in a single GPU. The approach aims to improve throughput and reduce latency for graph-based ANNS in storage-disaggregated environments.