Showing 66 papers for 2026-05-12
PathBoost introduces a gradient-boosting framework that learns discriminative path-based features directly from graphs for graph-level classification and regression. It extends prior work with a logistic loss for binary classification, a prefix-based decomposition to incorporate multiple node and edge attributes into the path feature space, and automatic anchor selection to improve feature construction.
Federated Graph Generalized Category Discovery (FGGCD) enables discovering novel categories across decentralized graph clients while preserving knowledge of known categories. The method targets the open-world setting in federated graph learning by tackling cross-client representation alignment and the incremental integration of new categories without sharing raw data.
Efficient Prompt Learning for Traffic Forecasting proposes a prompt-based adaptation framework for spatio-temporal graph neural networks to improve generalization under distribution shifts with reduced fine-tuning. By using trainable prompts, the model remains robust across varying traffic regimes and locations.
PRAETORIAN targets GNN backdoors by focusing on the intrinsic requirements of effective backdoors rather than surface cues. It observes that flipping a victim node's prediction requires substantial influence and that attackers tend to inject many trigger nodes, guiding robust defenses against adaptive attacks. The method analyzes internal correlations and external influences to detect and mitigate backdoors.
Graph Computation Meets Circuit Algebra argues that successful GNNs for electronic design automation should align their propagation, aggregation, and supervision with the native algebra of the task. For example, static timing analysis maps to max-plus/min-plus recurrences on DAGs, while placement relies on hypergraph wirelength and density penalties, guiding task-aware differentiable solvers.
GNN for Structural Displacement Prediction presents a data-driven framework that models structural systems as graphs with joints as nodes to predict displacements under external loading. It aims to replace or augment FEM with fast, real-time predictions for structural health monitoring, balancing accuracy and efficiency.
Structure-Centric Graph Foundation Models (SCGFM) treat topology as the primary transferable knowledge and model graphs as metric measure spaces with learnable geometric bases that define a shared structural coordinate system. Graphs are aligned to these bases via Gromov-Wasserstein distances, yielding structure-aligned latent representations that transfer across domains.
Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning proposes a structural pruning method for SO(3) equivariant GNNs to reduce computational cost while preserving accuracy in atomistic simulations. Pruning is performed along channels and tensor orders to achieve efficient, scalable models.
AutoSlim is a machine learning-based framework that prunes automata graphs for hardware accelerators by leveraging features from prior executions and a Random Forest classifier to identify low-impact nodes and edges. The data-driven pruning reduces memory usage and accelerates runtimes without sacrificing performance.
PACT (Peak-Aware Cross-Attention Graph Transformer) delivers an efficient storm-surge emulator by encoding each forcing patch as a graph with GraphSAGE and using a learned station query to aggregate information through cross-attention. The model achieves accurate, fast predictions suitable for ensemble hazard assessments.
Hierarchical Attention-based Heterogeneous GNN (HA-HeteroGNN) addresses interpretability and efficiency by a two-tier attention mechanism across 16 node types and 18 edge types, enabling per-node relevance-based pruning. The framework provides an explainability-to-pruning pipeline for scalable heterogeneous graph reasoning.
RAwR (Role-Aware Rewiring via Approximate Equitable Partition) augments inputs with a quotient graph from approximate equitable partition to alleviate oversquashing and enable long-range interactions. The rewiring improves information flow in graphs with bottlenecks, while maintaining scalable computation.
CTQWformer fuses continuous-time quantum walks (CTQW) with a Transformer by using a trainable Hamiltonian to blend graph topology and node features. This physically grounded approach models quantum walk dynamics to capture rich global dependencies for graph classification.
End-to-End Keyword Spotting on FPGA Using Graph Neural Networks with a Neuromorphic Auditory Sensor presents the first integrated FPGA implementation that combines a Neuromorphic Auditory Sensor (NAS) and a GNN for keyword spotting, enabling on-device processing with sparse event-based data.
ChaosNetBench provides a synthetic Chaos dynamics–based benchmark for spatio-temporal GNNs, enabling evaluation across controlled chaotic regimes. It supports comparisons of architectures under diverse dynamical conditions beyond real-world datasets.
UFO is a unified flow-oriented framework for robust continual graph learning that handles both catastrophic forgetting and noise in newly arriving graph portions. It integrates flow-based mechanisms to maintain stable learning on evolving graphs.
Learning Graph Foundation Models on Riemannian Graph-of-Graphs (R-GFM) treats structural scale as a first-class concept using a Riemannian GoG representation, enabling scale-aware transfer across graph domains and mitigating scale mismatch from fixed-hop subgraph sampling.
Anchor-guided Hypergraph Condensation with Dual-level Discrimination distills large hypergraphs into compact synthetic ones. It uses anchor-guided condensation with dual-level discrimination to preserve both structural fidelity and discriminative features, jointly optimizing the structure generator and condensed features.
One-Step Graph-Structured Neural Flows (GSNF) model irregular multivariate time series with a single mapping that captures inter-variable interactions. It addresses the limitations of treating variables independently by using graph-structured neural flows to model dependencies and improve classification.
Teaching LLMs to See Graphs: Unifying Text and Structural Reasoning introduces Graph Transformer Language Model (GTLM), a pretrained LLM architecture that natively processes graph topologies to alleviate the bottleneck of compressing graphs into single tokens. GTLM enables joint text and structure reasoning without heavy pipelines.
This work shows that principled relation operators in knowledge graph embeddings must satisfy linearity, trace preservation, and complete positivity, which together imply a Kraus channel structure via the Kraus representation theorem. The completeness constraint is shown to be equivalent to these axioms, offering a foundational alternative to externally imposed design choices. By recasting relations as Kraus channels, the authors unify and constrain common KG embedding operators while enabling principled generalization.
CMKL introduces modality-aware continual learning for evolving biomedical knowledge graphs. It leverages multimodal signals and targeted regularization to adapt to changing graph structures while avoiding uniform treatment across modalities. Experiments show improved accuracy and robustness over baselines on evolving KGs.
We propose topology-aware structural graph encoding that preserves chain-scale morphology in polymers, addressing the limits of using only repeat-unit graphs. The encoding captures how chain distribution influences properties and improves predictive performance under small data regimes. This leads to more accurate polymer property predictions.
We propose a general pre-training strategy for QSAR: train GNNs to predict Extended-Connectivity Fingerprints (ECFP). This pre-training improves downstream QSAR performance, particularly under out-of-distribution splits, with five of six Biogen benchmarks showing statistical significance.
We study in-context graph learning in large language models with a toy graph random-walk over two competing structures. The results indicate that neither global topology nor local transitions alone suffice. PCA of intermediate representations reveals both topologies are encoded in orthogonal subspaces at certain mixture ratios, suggesting a mixed mechanism.
This survey reviews attention-based GNNs, outlining how attention helps select discriminative features and filter noise. It provides a taxonomy, discusses design choices, datasets, evaluation practices, and practical guidance, and highlights gaps and directions for future work.
We address the dependence on temporal context by enabling structure-only initialization for GNN-based MD simulators. This init uses static structural information, enabling better out-of-distribution generalization while preserving differentiability for inverse design tasks. The results show improved generalization and potential benefits for design workflows.
PrimeKG-CL introduces a continual learning benchmark for evolving biomedical KGs, assembled from nine authoritative sources with realistic update patterns. It captures asynchronous growth, edge deprecation, and modality changes to stress CGL methods. The paper provides baselines and evaluation protocols.
We propose a unified framework that jointly models co-methylation graph structure with site-specific DNA sequence context to predict epigenetic age. This integration improves accuracy over methods using only sequence or graph information and supports more reliable age estimations.
We demonstrate end-to-end real-time deployment of a dynamic GNN for online event selection in a collider trigger, implemented on FPGA-based hardware (Versal VCK190). The work addresses latency and throughput constraints with hardware-software co-design and reports practical viability for big-science triggers.
We propose an explanation-based detection framework for backdoored GNNs that combines multiple explanation-derived metrics rather than relying on a single indicator. The approach improves robustness against diverse backdoor strategies and outperforms prior single-metric detectors.
We explore explainable node similarity by augmenting GNN-based similarity with explanations that attribute which substructures or features drive the similarity. The framework yields human-interpretable reasons for why nodes are considered similar.
We introduce Test-Time Structural Alignment (TSA), a Graph Test-Time Adaptation method that aligns a pretrained GNN to the target graph structure at test time, enabling adaptation without retraining. This yields better performance under distribution shifts.
This work discusses LLM-enhanced text-attributed graph learning and proposes a unified framework to systematize diverse optimization perspectives when using LLMs to generate graph descriptions. It also outlines key challenges and practical guidelines.
We propose Local-UID feature augmentation for GNNs solving ILPs within Learning-to-Optimize, avoiding issues with globally unique IDs. The approach preserves expressiveness while reducing spurious correlations, improving generalization.
GraphBench provides a comprehensive benchmark for graph learning with diverse datasets and tasks, standardized evaluation protocols, and support for graph foundation models. It aims to improve reproducibility and enable thorough evaluation beyond narrow benchmarks.
DUALFloodGNN integrates data-driven GNN predictions with physical constraints to improve realtime flood modeling. The physics-informed approach yields faster, more reliable forecasts useful for operational flood management.
We formulate adversarial attacks guided by Optimal Transport on GNN-based bot detectors, considering domain and temporal constraints. The results show attackers can degrade detection under realistic conditions, underscoring the need for robust defenses.
We perform controlled ablations of a minimal transformer modification with four components—sparse adjacency masking, edge-type biases, query scaling, and value gating—to isolate their contributions. We find sparse adjacency masking accounts for the majority of improvement in three-hop and other benchmarks, while the other components contribute less.
We show that pairwise message passing limits learning for multi-agent pathfinding. We propose Hypergraph Neural Networks to capture higher-order interactions among agents, and experiments show improved coordination and fewer collisions compared with pairwise GNN baselines.
This paper reframes text-rich graphs by treating node text as the primary medium through which structure is expressed, not just a static attribute. It argues that existing workflows compress text into embeddings before reasoning, creating an information bottleneck, and proposes the RAMP framework to rethink message passing to leverage raw text for graph reasoning.
Knowledge graphs are often incomplete, and textual descriptions offer rich information for entities. This work learns inductive entity representations from text to support link prediction and examines generalization to unseen entities.
The paper questions the assumption that neural CQA models generalize beyond explicit graph structure by comparing them with a training-free query relaxation strategy that relaxes constraints and counts paths. Across multiple datasets, the results illuminate when neural methods truly extrapolate and when simple relaxation suffices.
We introduce SKG-VLA for multimodal decision making by modeling scene structure and priors in a knowledge graph to integrate heterogeneous evidence (narratives, screenshots, metadata, policies). The approach leverages explicit scene structure and cross-evidence dependencies to improve decision quality.
EpiGraph constructs a large epilepsy knowledge graph and benchmark to support evidence-intensive clinical reasoning, integrating 48,166 peer-reviewed papers and seven clinical resources into 24,324 entities and 32,009 evidence-grounded triplets.
TESSERA is a 3-part neuro-symbolic framework that uses LLMs to provide local guidance for Monte Carlo Tree Search over knowledge graphs, enabling the composition of mechanistic explanations for drug–disease pairs.
HAGE introduces a weighted, multi-relational memory framework where memory retrieval is treated as sequential, query-conditioned traversal over a unified relational memory graph. This design allows varying relation strengths and confidence to influence retrieval for agentic LLM systems.
MAGE externalizes self-knowledge of multiple agents into a four-subgraph co-evolutionary knowledge graph, supporting what each agent learns and how it preserves knowledge across iterations. The framework enables self-evolving agents to coordinate and reason with a frozen backbone at inference.
SLASH investigates how LLMs internalize graph structure, finding a sawtooth pattern in attention that indicates spontaneous topology reconstruction, offering a path to sharpen internal structural attention without external adapters.
PathISE proposes learning informative path supervision for KGQA, reducing the need for costly intermediate annotations by guiding evidence retrieval along question-relevant KG paths and subgraphs, enabling better grounding of LLMs.
Crystal Fractional Graph Neural Network (GNN) for energy prediction in high-entropy alloys combines local crystal environments with global composition using graph attention to capture both intra-site interactions and overall alloy composition.
This paper presents a provenance-aware modular pipeline to convert historical handwritten tabular images into knowledge graphs, ensuring transparency, reproducibility, and trust by tracking each transformation step.
We propose a multi-level graph attention network with contrastive learning across multiple views to enhance knowledge-aware recommendations, addressing sparse labels, limited structure learning, and noisy KG entities.
ShadowMerge introduces a novel poisoning attack on graph-based agent memory by exploiting relation-channel conflicts, enabling an attacker to insert conflicting relations that later influence agent decisions.
HOME-KGQA provides a benchmark dataset for multimodal knowledge graph question answering focused on household daily activities, addressing bias toward encyclopedic knowledge by offering household-relevant scenarios.
Oracle Poisoning defines attacks where adversaries corrupt knowledge graphs that AI agents query at runtime, demonstrating six scenarios on a large production code KG to weaponize agent reasoning.
MicroWorld builds a multimodal attributed property graph from scientific image-caption corpora and uses it to augment multimodal LLM reasoning at inference, bridging the microscopic domain gap without additional training data.
GraphReAct extends reasoning-acting frameworks to graphs by interleaving evidence retrieval from graphs with multi-step reasoning and actions, enabling dynamic, iterative graph inference with LLMs.
The paper proposes watermarking GNNs via explanations to protect ownership, addressing limitations of backdoor-based methods and avoiding data poisoning while embedding verifiable ownership signals.
ArchRAG introduces an attributed community-based hierarchical retrieval-Augmented Generation framework to improve relevance and efficiency in graph-based QA by organizing retrieval around graph communities.
We propose TextBridgeGNN, a pre-trained graph neural network for cross-domain recommendation that uses text-guided transfer to bridge domain gaps. It addresses the non-transferability of ID embeddings across isolated domains and structural heterogeneity between heterogeneous interactions by aligning cross-domain semantics via textual information. This results in a more transferable recommender that can perform across domains and cold starts.
We introduce Response-G1, a framework for proactive streaming video understanding that builds explicit scene graphs to align accumulated visual evidence with the query's expected response conditions. It performs three fine-tuning-free stages, including online query-guided scene graph generation from streaming clips and explicit alignment between evidence and response needs. This structured approach improves when-to-respond decisions during video unfolding.
GraphInstruct presents a progressive benchmark to diagnose capability gaps in LLM-based graph generation. It evaluates instruction-following graph synthesis across diverse graph types and tasks, revealing where models struggle and guiding targeted improvements. It advances beyond pre-LLM graph-generation reviews by focusing on instruction adherence and practical prompting.
GraphNetz is a benchmarking framework for GNNs that emphasizes statistical reporting over raw accuracy tables. It provides confidence intervals, paired tests, multiple-comparison corrections, and rank aggregation to yield structured statistical comparisons across datasets, seeds, and models. The default output is a structured statistical report rather than a plain accuracy table.
Global-local Spatial-temporal Aware Graph Attention Network for Network Traffic Forecasting proposes a model that addresses limitations of predefined spatial adjacency by capturing global and local spatial dependencies and dynamic temporal patterns with graph attention. It yields improved accuracy and robustness in forecasting network traffic.
Toward Multi-Database Query Reasoning for Text2Cypher discusses extending Text2Cypher from a single graph database to a multi-database setting, enabling queries across several independent graphs with varying schemas. It introduces mechanisms for database selection, schema alignment, and cross-graph Cypher query composition to enable seamless cross-domain data access.