Showing 14 papers for 2026-06-24
We introduce a deep learning approach that reconstructs monthly GRACE-like terrestrial water storage anomalies (TWSA) back to 1940 by learning the relationship between daily observations and TWS changes. The method uses spatio-temporal graph neural networks to capture complex land-atmosphere interactions and is demonstrated for South America.
This work provides a fair evaluation of Graph Foundation Models for node property prediction, comparing different GFMs across diverse tasks and datasets. It discusses evaluation protocols, potential biases, and proposes benchmarks to ensure fair comparisons.
We propose a benchmarking methodology for GNNs based on the graph alignment problem, a combinatorial task that aligns two unlabeled graphs to maximize edge overlap. We generate datasets with synthetic and real-world graphs, creating a family of alignment tasks with increasing difficulty to study GNNs and positional encodings.
FuseSampleAgg introduces a one-pass neighborhood estimation technique for updating knowledge-graph embeddings under tight time, memory, and audit budgets. It replaces costly neighbor-sampling pipelines with a single-pass estimator that avoids subgraph materialization and reduces kernel launches, enabling efficient KG refresh and validation.
We formulate optimization problems to generate adversarial inputs that induce large errors between a neural network's AC power-flow predictions and the true AC power-flow equations. Demonstrated on the CANOS-PF GNN with a 14-bus test grid, the adversarial points yield substantial errors in reactive power and voltage magnitude, highlighting system vulnerabilities.
SEAGAN is a domain-specific and edge-aware Graph Attention Network designed for dynamic plant processes, modeling interdependent processes such as the A-Ci curve that relates photosynthesis to intercellular CO2. By incorporating domain constraints and edge-level information, SEAGAN improves estimation of photosynthetic parameters.
Macro Graph of Experts (MGOE) is the first framework to leverage macro graph embeddings for billion-scale multi-task recommendation, capturing task-specific relationships across massive graphs. By sharing information through macro graphs, MGOE improves performance and scalability for diverse tasks.
T2D-Bench is an evidence-gated benchmark for evaluating LLM outputs in Type 2 diabetes care, checking whether model responses satisfy explicit, graph-checkable evidence requirements. Built on a multi-layer clinical-lifestyle knowledge graph that combines biomedical spines, DrugBank, ADA standards, and lifestyle rules.
We study federated long-tailed graph learning and propose an energy-guided dual decoupling approach to balance representation and classifier learning under data scarcity and distributed settings. The method aims to improve performance on minority classes while respecting privacy constraints.
Subjective-Graph LLM Agents study how graph-local evidence and credibility-weighted communication among agents with individualized subjective graphs can produce persistent distortions in perceived academic standing despite objective signals. A data-constrained multi-agent framework enables agents to form judgments through their own evidence networks.
Rule2Text presents a framework that uses LLMs to generate natural language explanations for mined knowledge-graph rules and to evaluate their quality. The approach is tested across multiple KG datasets, including variants of Freebase.
We propose a training-free mixture-of-agents framework for multi-document summarization that combines LLMs with knowledge graphs to model inter-document relations without supervised training. The framework deploys diverse agents to extract, fuse, and generate summaries.
We address event-stream recognition by combining point-voxel downsampling with graph neural networks; traditional max/mean pooling pools node embeddings, which loses information. We propose a point-voxel absorbing graph representation capable of preserving temporal/spatial structure, achieving better recognition.
Knowledge-Graph Grounding Helps LLMs Only for Out-of-Training Knowledge: A Controlled Study on Clinical Question Answering investigates when KG grounding improves LLM performance, reproducing previous results and finding that grounding mainly helps for knowledge outside the training set; for in-domain knowledge, grounding offers limited gains.