Showing 20 papers for 2026-06-17
This work develops an end-to-end graph neural surrogate to forecast CO2 plume migration in complex geological formations, enabling fast, data-driven predictions of multiphase flow. The approach is evaluated on the SPE11A benchmark, tackling sharp gas–water interfaces, strong advection, and fingering, demonstrating accuracy and speed advantages over traditional simulators.
The paper proposes using the Finsler Laplacian as a nonlinear generalization of the Laplace-Beltrami operator for graph-based learning on point clouds. They prove convergence of the discrete Finsler estimates to the true operator as sample size grows, and show it can be implemented as a graph neural network layer, enabling anisotropic geometric learning.
This work shows that adding LLM-generated features to GNNs by simple input concatenation can harm performance on standard homophilous graph benchmarks, unlike end-to-end LLM pipelines. Using an MLP backbone with Planetoid splits, pure concatenation of SBERT/GPT-derived features reduces accuracy, highlighting interference rather than benefit.
The paper tackles feature heterogeneity across graphs for universal pre-trained graph models by avoiding reliance on textual data. They introduce learnable graph patches as the smallest semantic units, enabling flexible, dataset-agnostic representations and better transfer across graphs.
The paper analyzes knowledge graph foundation models and how they generalize across unseen graphs. It reveals that half-links (partially observed triples) can be surprisingly informative, such that predicting a test triple may require observing only one end, explaining nonuniform generalization.
This work improves graph-based network intrusion detection by making the model sensitive to temporal evolution of traffic. It introduces timestamp-aware spatio-temporal graph contrastive learning to capture changing attack patterns and improve generalization to unseen intrusions.
The paper investigates fairness in diffusion-based graph neural networks and identifies biases that can propagate or amplify data biases. It proposes a fairness-aware adaptation of diffusion-based GNNs to mitigate such biases while maintaining performance.
We introduce Manifold GCN for graphs with features on a Riemannian manifold, proposing a diffusion-based layer and a tangent MLP that operate on manifold-valued data. Both layers are permutation and isometry-equivariant, providing an inductive bias that respects the geometry.
This work introduces SMGFM, a spectral pretraining framework for multimodal-attributed graphs that separates structure-induced semantics from modality-intrinsic semantics. It aims to align topology with multimodal node content (text, images, etc.) to improve downstream tasks.
The paper tackles out-of-distribution blockchain anomaly detection by incorporating temporal motif awareness into graph test-time adaptation. The framework addresses adversarial pattern evolution and semantics shift, enabling robust detection on unseen blockchains.
FlowRAG introduces explicit reasoning over a frequency-aware, multi-granularity graph flow to improve retrieval-augmented generation. It aims to reduce under-retrieval and brittle multi-hop reasoning that can derail entity-to-entity transitions.
The paper analyzes the logical expressiveness of graph neural networks from a semantic viewpoint, showing how logical formulae correspond to restricted GNN architectures and how GNNs can be translated into logical representations. It clarifies the structural preservations that govern expressive power.
This work proposes a unified framework for context-aware and relation-aware retrieval-augmented generation on graphs, addressing limitations of entity-centric and chunk-centric methods that retrieve information separately. It fuses retrieved content to enable more coherent and knowledgeable generation.
The study applies graph neural networks to fuse cybersecurity and drone intelligence in a cyber-physical setting, focusing on intrusion detection and UAV management. It provides an integrated procedure for using GNNs to understand network structures and identify malicious activity.
This work introduces a trust-aware coordination framework for multi-agent software artifact management, where confidence estimates propagate through a shared knowledge graph to prevent error propagation. It focuses on improving traceability and consistency across sequential AI-driven tasks.
The paper investigates semi-supervised image classification with graph neural networks that aggregate multiple feature types from various extractors. It discusses the benefits of combining diverse features (e.g., CNN, ViT) for improved performance and robustness.
This paper proposes improved negative sampling strategies to train knowledge graph foundation models, addressing the limitations of random negatives. By selecting more informative negatives, the approach enhances zero-shot generalization to unseen KGs.
This work applies a Graph Matching Network for Alzheimer's disease diagnosis and introduces test-time domain adaptation to handle inter-site heterogeneity in multi-centered MRI data. The method improves diagnostic accuracy across datasets.
The paper analyzes genealogical networks at scale to identify universal structural properties, combining network-theoretic structure with an inferred temporal dimension. Using over a hundred datasets, it reveals how kinship networks exhibit characteristic structural and temporal hallmarks.
This work investigates whether generative recommenders deepen information cocoons by simulating user interactions with LLM-powered user simulators in a closed-loop setup. It studies how replacing atomic item IDs with Semantic IDs affects exposure diversity and bias, evaluating cocoons.