Showing 17 papers for 2026-03-04
This work proposes a hypergraph neural network approach to predict network controllability robustness (NCR) under various attacks, aiming to replace time-consuming attack simulations with fast predictions on large networks. By modeling high-order relations via hypergraphs, it captures complex interdependencies that traditional graph models miss, enabling scalable NCR estimation and robustness analysis.
NETRA introduces a Transformer-based multimodal graph framework that prioritizes disease genes via attention-driven relevance scores rather than static centrality metrics. It integrates gene regulatory networks across modalities to better capture cross-modal biological heterogeneity in Alzheimer's disease.
MASPOB develops a bandit-based prompt optimization method for multi-agent systems that use graph neural networks. It addresses sample efficiency and the sensitivity of MAS performance to prompts, by learning prompts online through exploration-exploitation to boost coordination without changing workflows.
This work tackles zero-shot graph learning with large language models by adaptive subgraph denoising and a text-based graph reasoning paradigm (Graph-R1) that reduces cross-modal alignment issues. It relies on LLMs to predict graph tasks in a purely textual format.
We propose Multi-scale Adaptive Neighborhood Aware Transformer for Graph Fraud Detection, a transformer model that uses multi-scale neighborhoods to detect fraud signals and overcome GNN biases like homophily and limited global modeling. It improves detection by capturing both local and broader neighborhood information.
VL-KGE presents a framework marrying vision-language models with knowledge graph embeddings to handle heterogeneous multimodal knowledge graphs. It addresses cross-modal alignment and modality availability issues, enabling richer multimodal reasoning over KGs.
ChemFlow is a hierarchical neural network for multiscale representation learning in chemical mixtures. It models interactions from atoms through functional groups to whole molecules while accounting for mixture composition, enabling accurate property prediction.
Graph Homomorphism Distortion introduces a graph-level (pseudo-)metric based on graph homomorphisms to distinguish graphs with features and to evaluate latent representations. The metric measures worst-case distortion under homomorphisms, helping compare structure-features interplay and latent spaces.
This paper develops quantized SO(3)-equivariant GNNs to enable efficient molecular property prediction on edge devices. They propose magnitude-direction decoupled quantization and branch-separated quantization for vector features to preserve equivariance while speeding up inference.
EP-GAT is an energy-based parallel graph attention network for stock trend classification. It models dynamic inter-stock dependencies and preserves hierarchical features within stocks, offering improved forecasting over traditional static GNN approaches.
Odin is a production-grade graph intelligence engine for autonomous discovery in knowledge graphs. It guides exploration with the COMPASS score, combining structural importance via Personalized PageRank and semantic plausibility via Neural Probabilistic Logic Learning as a discriminative filter.
The paper benchmarks graph and non-graph methods for Caenorhabditis elegans neuron classification using functional connectomes. Four GNNs and four baselines are compared across Spatial, Connection, and Neuronal Activity features; attention-based GNNs excel on Spatial and Connection features, while Activity features are less predictive.
The work presents a directed graph model and experimental framework to study time-dependent text visualization. It focuses on how evolving textual relationships (e.g., news articles) change over time and how visualizations should support interpretable understanding of narratives.
MIRAGE introduces a knowledge-graph-guided cross-cohort MRI synthesis framework to support Alzheimer's disease prediction when MRI is missing. By leveraging KG-guided synthesis of 3D scans from tabular data, it addresses modality missingness and aims to improve cross-cohort AD prediction.
The paper proposes a law article recommendation method for Chinese criminal law by combining a Case-Enhanced Law Article Knowledge Graph (CLAKG) with a Large Language Model. The CLAKG stores current articles and historical cases to enable efficient retrieval and accurate recommendations.
Efficient Maintenance of Leiden Communities in Large Dynamic Graphs presents methods to update Leiden communities efficiently as graphs evolve, avoiding full recomputation. The approach provides detailed update procedures to maintain accurate communities in dynamic, large-scale graphs.
The work studies link prediction for event logs in the process industry to support retrieval-augmented generation (RAG) systems. It addresses fragmented shift-book records and proposes methods to predict related event-log links to improve knowledge management and data retrieval.