Showing 27 papers for 2026-04-28
We present BiTA, a Bidirectional Gated Recurrent Unit–Transformer aggregator within a Temporal Graph Network framework for proactive alert prediction in computer networks. By enabling bidirectional, multi-scale temporal aggregation, BiTA captures recursive attack patterns that unidirectional TGNs miss. On network alert datasets, BiTA improves predictive performance against strong baselines.
This paper investigates the expressive power of first-order properties expressible by ACR-GNNs under global readout. It proves that sum aggregation combined with a global readout suffices to capture FO properties that cannot be expressed in C2 logic on both directed and undirected graphs. This strengthens previous results and clarifies the capabilities of common GNN readouts.
The authors introduce HBGSA, a Hydrogen Bond Graph with Self-Attention for drug-target binding affinity prediction. The model builds a hydrogen-bond graph and applies self-attention to capture geometric constraints and interactions between ligand and target. It addresses limitations of sequence-based and traditional structure-based approaches, achieving improved binding affinity predictions for virtual screening.
We propose Layer Embedding Deep Fusion Graph Neural Network, which mitigates issues in deep GNNs under low-homophily settings by fusing layer embeddings across layers. This approach reduces structural noise from heterophilic edges and helps capture long-range dependencies in graphs. It yields improved performance on benchmarks with low homophily.
We introduce Hamiltonian Graph Inference Network (HGIN), which jointly recovers the interaction graph and predicts node dynamics for lattice Hamiltonian systems from trajectory data. Unlike prior methods that assume a given graph or homogeneous dynamics, HGIN learns both topology and dynamics in a unified framework. It handles nonseparable Hamiltonians and heterogeneity across nodes.
Crystal structure prediction using graph neural combinatorial optimization reframes CSP as a learnable combinatorial optimization on a discretized lattice guided by graph neural networks. The method searches for atom placements that minimize interaction energy with a learned optimization component. It demonstrates improved accuracy over traditional CSP baselines.
We propose model-agnostic explanations for temporal graph predictions using Shapley values. The work includes event-level (edge-level) explainers that apply KernelSHAP to temporal events and uses Owen values to attribute contributions across edges and time. These explanations reveal how temporal events shape TGNN predictions.
CMGL introduces Confidence-guided Multi-omics Graph Learning for cancer subtype classification. It first estimates per-sample modality reliability with evidential deep learning, then freezes confidence estimates to weight modalities in the graph learning pipeline. This yields more robust subtype predictions by mitigating noisy omics data.
Latent-Hysteresis Graph ODEs (HGODE) couple feature evolution with a latent hysteresis mechanism to model coupled topology-feature evolution via continuous phase transitions. This design overcomes the monostability trap of Graph ODEs with positive irreducible mixing and allows richer dynamics. Empirical results show HGODE better captures phase-transition-like behavior.
PathMoG introduces a pathway-centric modular-GNN for multi-omics survival prediction. It reorganizes genome-scale inputs into 354 KEGG-informed pathway modules and uses a Hierarchical Omics Modulation module to condition gene representations on mutation, copy-number variation, pathway context, and clinical data, followed by dual-level pooling for survival risk. The approach improves predictive performance.
This paper develops spatio-temporal graph neural networks for fraud detection in cryptocurrency markets. It captures coordinated manipulations across assets and their transactional patterns, modeling both spatial (asset relations) and temporal dynamics. The proposed model outperforms asset-centric baselines in detecting market fraud.
We propose a probabilistic graphical model using graph neural networks for Bayesian inversion of discrete structural component states. The approach handles ill-posed inverse problems by learning probabilistic mappings with GNNs to approximate posterior distributions over discrete states given measurements. The method is demonstrated on civil infrastructure data.
ComplianceNLP is an end-to-end system that monitors regulatory changes, extracts structured obligations, and detects compliance gaps against institutional policies. It uses a knowledge-graph-augmented retrieval-augmented generation pipeline to ground generations in regulatory facts. The system demonstrates effective gap detection across multi-framework regulations.
XGRAG presents a graph-native framework for explaining KG-based retrieval-augmented generation. It moves beyond black-box RAG by interpreting how specific knowledge graph elements influence the LLM output. The method provides graph-based explanations for KG-RAG reasoning.
Question-Adaptive Graph Learning for multi-hop RAG adapts the retrieval graph to the question context to identify multiple knowledge targets. It improves multi-hop retrieval quality by reducing noise and guiding the model to relevant evidence. Experimental results show gains over static-graph baselines.
The paper shows that verification of quantized ACR-GNNs with global readout is decidable but highly intractable. It defines a logical language for reasoning about quantized GNNs and proves co-NEXPTIME-completeness, indicating extreme computational hardness. Experiments suggest quantized GNNs remain lightweight in practice.
BEAR proposes beam-search-aware optimization for LLM-based recommendations. It identifies a training-inference mismatch where SFT optimizes overall likelihood but not beam-ranked retrieval. The proposed objective aligns training with beam search, improving recommendation performance.
Mochi is a Graph Foundation Model that uses meta-learning to align pre-training and inference for efficiency and task unification. It critiques reconstruction-based pretraining and uses meta-learning to bridge pretraining tasks with downstream requirements, improving generalization and efficiency.
Distance-Misaligned Training studies when global graph transformers misallocate communication relative to the signal locality. Using a synthetic contextual stochastic block model, the authors quantify misalignment between where the label-relevant information lies and where the model aggregates information. They propose strategies to adapt training to the appropriate distance scales.
Graph-to-Vision investigates using vision-language models to interpret and reason about visualized graphs. It introduces a benchmark for multi-graph understanding and reasoning, evaluating how VLMs handle cross-graph reasoning tasks in graph data.
We propose Eidolon, a post-quantum signature scheme grounded on the NP-complete k-colorability problem (k ≥ 3). The construction generalizes the Goldreich–Micali–Wigderson zero-knowledge protocol, applies the Fiat-Shamir transform, and uses Merkle-tree commitments to compress signatures from O(tn) to O(t log n). Hard instances are generated by planting a coloring while preserving the statistical profile of random graphs, with empirical security analysis against classical solvers (e.g., ILP, DS).
We introduce Causal Concept Graphs (CCG): a directed acyclic graph over sparse, interpretable latent features in LLMs where edges capture learned causal dependencies between concepts. The approach combines task-conditioned sparse autoencoders for concept discovery with DAGMA-style differentiable structure learning to recover the graph, and introduces the Causal Fidelity Score (CFS) to evaluate whether graph-guided interventions induce the intended causal effects.
This paper summarizes the LLM+Graph Workshop at VLDB 2025, focusing on integrating large language models with graph-structured data through algorithms and systems spanning graph data management and graph machine learning for real-world applications. It highlights key research directions, challenges, and innovative solutions presented at the event.
The study explores calibrating pipe roughness in large-scale water distribution networks by partitioning the network using hydraulic and graph-derived attributes to improve calibration stability. Using a real-network high-fidelity model as a benchmark, it evaluates density-based clustering and topology-driven grouping; optimization experiments show attribute-based grouping yields stable, repeatable results comparable to traditional methods.
Behavioral Intelligence Platforms advocate moving from query-driven analytics to active systems that continuously detect and explain behavioral phenomena, using probabilistic journey graphs, behavioral knowledge extraction, and grounded language generation to produce autonomous insights.
The paper provides a unified framework to analyze graph-based Retrieval-Augmented Generation (RAG) methods under a common perspective, synthesizing how they incorporate graph data with LLMs and external knowledge sources. It systematically compares methods under standardized settings, outlining coverage, limitations, and future directions.
Time Agnostic Library (TAL) enables live temporal queries over RDF change histories: a Python library that performs temporal SPARQL queries on any SPARQL-compliant triplestore, supporting six identified temporal retrieval needs and concurrent updates without offline ingestion.