Showing 33 papers for 2026-03-31
We study GNN-based entity resolution in master data management by modeling it as a bipartite graph between entities and attribute values. The work develops a tight expressivity hierarchy to identify the cheapest MPNN architecture that provably works for a given matching criterion, along with theoretical guarantees. It provides clear conclusions on the minimal architectures required for different matching tasks.
We present GSR-GNN, a Grouped-Sparse-Reversible GNN designed for circuit graphs to enable deep architectures. The framework reduces both compute and memory overhead through grouping, sparsity, and reversible layers, allowing hundreds of layers to be trained feasibly. Experiments show deep GNNs outperform shallow counterparts on circuit-analysis tasks under resource constraints.
We introduce Cross-attentive Cohesive Subgraph Embedding to mitigate oversquashing in GNNs. The method enriches node representations by aggregating information from cross-attentive cohesive subgraphs, improving long-range information propagation. It yields performance gains particularly in dense and heterophilic regions on relevant benchmarks.
CrossHGL is proposed as a text-free foundation model for cross-domain heterogeneous graph learning. It enables cross-domain generalization without requiring shared schemas or textual attributes by learning domain-agnostic representations. Empirical results demonstrate strong transfer capability across domains.
TMTE tackles the quality of topology and modality in multimodal attributed graphs by introducing Task-aware Modality and Topology Evolution. It jointly optimizes modality features and graph structure in a task-aware manner. Experiments show that this co-evolution improves performance over single-graph baselines.
FedDES introduces a graph-based dynamic ensemble selection framework for personalized federated learning. Clients form a graph of peer relationships and selectively aggregate from the most beneficial peers to achieve instance-level personalization. It reduces negative transfer and improves accuracy and convergence compared to prior pFL approaches.
Graph Vector Field (GVF) models health risk as a vector-valued field on time-varying simplicial complexes, coupling discrete differential-geometric operators with modality-structured mixture-of-experts. Risk fields evolve using discrete operators like the Hodge Laplacian to capture cross-modal dynamics. The framework demonstrates effectiveness on wearable and environmental data.
ORACAL provides a robust and explainable multimodal framework for smart contract vulnerability detection with causal graph enrichment. Causal graphs capture control/data dependencies while multimodal cues are integrated to improve detection and provide explanations. The approach demonstrates resilience against adversarial perturbations and offers interpretable evidence.
NeiGAD augments graph anomaly detection by explicitly modeling spectral neighbor information. It integrates spectral features of neighbors and their attribute coherence into anomaly scoring in a plug-and-play manner. Empirical results show improved detection across standard GAD benchmarks.
FairGC introduces fairness-aware graph condensation that preserves fairness constraints while compressing data. The synthetic graphs avoid amplifying demographic disparities and improve the fairness-utility tradeoffs. Demonstrations on sensitive applications like credit scoring and social recommendation illustrate its benefits.
We propose contextual graph representations for task-driven 3D perception and planning in robotics. The approach extracts task-relevant object relations to reduce the state space and support efficient planning. Experiments show notable gains in perception accuracy and planning efficiency.
We present GATGraphClassifier, an Attention-based GNN for autism detection using fMRI-based functional connectivity from ABIDE. Functional connectivity matrices are constructed with Pearson correlation and fed into a graph attention network for classification. The model achieves higher ASD detection accuracy than baselines.
Continual Graph Learning: A Survey reviews continual learning for graphs, including replay-based and generative replay methods, and discusses condensation challenges. It analyzes limitations of distribution-matching approaches and outlines a taxonomy and future directions. The survey provides a comprehensive guide for researchers.
We introduce City-Networks, a large-scale dataset and a measurement framework to quantify long-range interactions in graph machine learning. The work compares models with global attention against local aggregation in capturing distant dependencies. It provides principled evaluation tools for long-range dependency capabilities.
Compact Conformal Subgraphs introduces graph-based conformal compression to construct smallest subgraphs that preserve predictive validity. It formulates compression as a weighted densest k-subgraph problem and aims to maintain conformal prediction guarantees with reduced structure. This benefits routing, planning, and sequential tasks.
Graph-Aware Stealthy Poison-Text Backdoors (TAGBD) study backdoor attacks on text-attributed graphs where attackers modify only node text while leaving the graph structure intact. The technique hides triggers in text to mislead downstream predictions. The work highlights security risks and discusses potential defenses.
PEANUT presents a gradient-free, restricted black-box attack on GNNs by injecting virtual nodes to disrupt eigenvector alignment in topology-driven message passing. The approach operates without gradients yet remains effective, exposing vulnerabilities of spectral GNNs. It emphasizes robustness concerns in topology-aware attacks.
LSM-GNN is a storage-based, multi-GPU training framework for large-scale GNNs that optimizes data transfer to improve scalability. It reduces memory footprint and communication overhead via storage-aware data management. Empirical results show substantial speedups and scalability improvements on large graphs.
A Survey of Heterogeneous Graph Neural Networks for Cybersecurity Anomaly Detection reviews HGNN approaches in cybersecurity, focusing on heterogeneity and temporal evolution. It provides a taxonomy, datasets, methods, and challenges with future directions. The survey serves as a comprehensive guide for applying HGNNs in cybersecurity.
AlignOPT aligns LLMs with graph neural solvers to improve combinatorial optimization by combining language model reasoning with graph-based solution methods. This approach enhances relational understanding and generalization to new instances. Experiments show improved performance and generalization on COP tasks.
GAAMA proposes Graph Augmented Associative Memory for agents to maintain persistent long-term memory across sessions. It aims to preserve structural relationships between memories, addressing limitations of flat retrieval-augmented generation and memory compression. The approach targets coherent, personalized agent behavior by reducing hub-dominated retrieval and enabling hierarchical reasoning.
GEAKG introduces Generative Executable Algorithm Knowledge Graphs, a class of knowledge graphs whose nodes encode procedural knowledge for algorithm design and operator composition. These graphs are executable and learnable, capturing problem-solving know-how and supporting reuse across domains.
This work presents a multi-view Graph Convolutional Network that fully leverages cross-view consistency. It overcomes limitations of KNN-based topology construction by avoiding fixed k-value choices, and it enhances inter-view feature consistency through a granular-ball topology and interactive fusion. The approach aims to improve multi-view representation learning with stronger consistency exploitation.
We propose DGDA, a Dual-branch Graph Domain Adaptation framework for cross-scenario multimodal emotion recognition. It addresses mismatches across conversation scenarios by aligning graph-structured representations of text, audio, and visual cues between source and target domains. The method improves transferability of MER models to unseen scenarios.
Amalgam combines probabilistic graphical models and large language models to synthesize data with both accuracy and realism. PGMs provide structured distributions but limited schemas, while LLMs offer rich schemas but skewed distributions. The hybrid algorithm balances these strengths to produce synthetic datasets suitable for advanced analytics, especially in healthcare.
Codebase-Memory constructs a persistent Tree-Sitter-based knowledge graph for LLM code exploration via the Model Context Protocol. It supports parsing 66 languages, uses a multi-phase pipeline with parallel workers, and provides call-graph traversal, impact analysis, and community discovery. The system enables structured, scalable codebase understanding for coding agents.
EvidenceNet offers a framework and dataset to build disease-specific knowledge graphs from full-text biomedical literature. An LLM-assisted pipeline extracts experimentally grounded findings as structured evidence nodes, normalizes entities, scores evidence quality, and links records by evidence type. This enables reasoning over provenance-rich biomedical evidence.
GammaZero introduces an uncertainty-aware graph representation to guide belief-space search in POMDPs. It provides a unified graph-based belief state across problem sizes, transforming belief states into graphs whose structural patterns learned on small problems generalize to larger ones. This enables domain-agnostic, scalable planning with neural generalization.
KG-Hopper enables compact open LLMs to perform knowledge-graph reasoning via reinforcement learning. It addresses knowledge-intensive reasoning bottlenecks in KBQA by guiding multi-hop KG reasoning with a learnable RL framework, mitigating error cascades of pipeline approaches and improving QA performance.
SRAG extends Retrieval-Augmented Generation by incorporating structured data into the retrieval step. By augmenting vector-based retrieval with structured data, SRAG improves retrieval relevance and factual grounding, resulting in more accurate LLM outputs.
RCLRec proposes Reverse Curriculum Learning to better model sparse conversions in generative recommendation. By starting with difficult conversion signals or reversing the usual curriculum, the method helps the model focus on hard-to-predict conversions and improves performance on sparse objectives.
Ontology-Compliant Knowledge Graphs study how to build KGs that are compliant with ontologies on internal and external levels. The paper introduces term-matching algorithms, pattern-based compliance, and new compliance metrics, using the building sector as a case study to validate the approach.
Link Prediction for Event Logs in the Process Industry addresses preprocessing for RAG by predicting links to complete fragmented event logs from shift books. The work tackles fragmented records in process industry data to improve graph-based retrieval and knowledge management applications.