Showing 14 papers for 2026-07-01
We propose relational and sequential conformal inference for energy time series over graphs using foundation models. Our approach combines spatial-temporal graph neural networks with conformal prediction to produce calibrated uncertainty intervals for energy demand forecasts, aiding risk-aware decision-making, grid stability, and planning in real-world energy systems.
FedLAB is a federated multimodal graph foundation learning framework that learns reusable semantic codebooks across distributed clients without sharing raw contents. It enables traceable, privacy-preserving knowledge transfer by aligning modalities and preserving intrinsic semantics in a federated setting.
TAG-DLM introduces a diffusion language model for text-attributed graphs that unifies textual reasoning and graph message passing. By masking and diffusing across both text and topology, the model performs joint reasoning over node descriptions and graph structure, overcoming the limitations of treating text and graphs separately.
Neighborhood-contextualized message passing enhances graph representations by incorporating broader local neighborhood context into each message. This extends standard pairwise center-neighbor messages and improves the model's ability to reason about complex relations in relational data.
DeXposure-FM is the first time-series, graph foundation model for measuring and forecasting inter-protocol credit exposures and systemic stability in decentralized financial networks. It combines temporal dynamics with graph structure to quantify contagion risk and forecast shocks, supporting risk management and policy design in DeFi.
Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations proposes using a typed knowledge graph as a grounding substrate to improve accuracy of LLM-based industrial asset operations. Questions are routed to KG-based reasoning or LLMs, e.g., via Cypher queries, yielding more reliable maintenance guidance across scenarios.
MedKGTab injects knowledge graphs into tabular medical data to enable cross-domain feature expansion. It infers uncollected biomedical features by exploiting statistical dependencies and medical correlations, using a row-column dual-attention mechanism to fuse information and expand features.
We propose a Gated Multi-Graph Fusion framework using Graph Attention Networks for Alzheimer's Disease detection from spontaneous speech. By transcribing audio with ASR to build semantic, dependency, and co-occurrence graphs, and fusing them through gating, the model captures content-structure-flow cues for detection.
Presents CVE-TTP KG, a knowledge graph linking software vulnerabilities to attacker behaviors from MITRE ATT&CK, bridging CVE/NVD data with tactics and techniques. This integrated graph enables better threat interpretation and faster response by revealing patterns of vulnerability exploitation across attack workflows.
Proposes a semantic-AI framework that automates cause-and-effect specification by coupling a knowledge graph with a constrained LLM layer, using a modular alignment ontology to represent process structure, operating modes, and fault scenarios. This enables consistent, scalable C&E generation for safety-critical processes.
Proposes homophily-aware compression for text-attributed graphs to improve LLM reasoning, preserving structural and semantic context while reducing context window requirements. The method uses graph homophily-guided sampling to reduce noise and stabilize reasoning in TAGs.
From a semantic perspective, we study the logical expressiveness of GNNs, linking fixed architectural choices to fragments of first-order and fixed-point logics. We show how certain aggregation, combination, and activation choices constrain what can be expressed, providing a semantic characterization of GNN capabilities.
Investigates graph-to-graph semantic similarity in knowledge graphs by evaluating KG embeddings for graph-level semantics. It argues that most embeddings focus on entities, relations, and triples, and proposes metrics and methods to compare whole graphs, highlighting limitations of existing approaches.
Extends beyond triplet plausibility by introducing relation set completion (RSC) for knowledge graphs, addressing incompleteness of entity-relation compatibility information. The paper defines the RSC task and discusses methodologies to complete relation sets for given entities, improving KGC beyond single triplet prediction.