Showing 17 papers for 2026-08-06
This work addresses accurate traffic forecasting for edge computing. It notes that traffic exhibits strong spatial correlations among neighboring regions and long-range temporal dependencies driven by mobility and application behavior. To handle non-stationary, long-horizon dynamics that challenge recurrent models, the paper proposes a Spatiotemporal Graph Transformer that captures both spatial and temporal structure for improved forecasting.
BBBP prediction with limited and imbalanced data; uses pre-trained molecular GNNs and geometry-informed parameter-efficient fine-tuning to adapt them to BBB permeability tasks, improving efficiency and performance.
EvtGraph introduces event-adaptive compression to align computation with temporal salience in multimodal time series. It reparameterizes sequences into event-level tokens via EAMC, selects a compact subset using a node budget, and performs temporally constrained sparse graph reasoning, enabling efficient processing of sparse data.
Hypergraph neural networks are vulnerable to label noise. This work conducts a systematic study of HGNNs under noisy labels and adapts representative label-noise learning methods (LLN) and graph-learning-with-noise (GLN) methods to hypergraphs to improve robustness for node classification.
Proposes a training-free geometric prior for neuronal representations based on tropical algebraic geometry. Using tropical Abel–Jacobi transform and polarization distances, it derives a descriptor that captures both topology and spatial geometry of tree-structured neuronal data for graph learning, mitigating WL limitations.
MACH is a modular framework for multimodal intent understanding that accounts for agreement and disagreement across modalities. It builds a hierarchical prototype hypergraph to encode both shared information and modality-specific conflict, improving interpretability and accuracy.
PriDyG proposes privacy-preserving dynamic graph inference by combining GNN-based structural learning with LLM-based semantic reasoning. It introduces edge-level differential privacy with incremental private multi-hop aggregation, buffering new edges and processing each edge exactly once, with parallel composition to bound privacy loss.
iStructTab introduces a structured feature sequencing approach for image-and-tabular multimodal learning. It uses GEDS, a sequencing method inspired by the Column Permutation Problem, to refine feature descriptors via similarity graphs and produce a better-aligned input for downstream multimodal models.
MT-GNN predicts brain morphometry by modeling mesh evolution in continuous time. A single graph network predicts the future per-vertex metric tensor (the first fundamental form) from arbitrary historical history and arbitrary forecast horizon, yielding a geometry-aware trajectory rather than only deformations.
CountTRuCoLa offers interpretable temporal KG forecasting based on symbolic rules. It learns four simple rule types, including temporal rules with confidence that combine recency and frequency; evaluation on nine datasets shows competitive performance with most state-of-the-art models, while remaining directly traceable to the rules.
Foundations of Equivariant Deep Learning unify graph and sheaf neural networks. The paper develops order-equivariant neural networks (OENN) that generalize standard message-passing and sheaf nets under a theory of equivariant vector bundles, extending symmetry-aware architectures beyond usual group actions.
MatrAIx presents a population-scale simulated-user evaluation platform for AI systems. It introduces Persona 8B with 8.3 billion persona records across 1,290 dimensions, generated from a dependency graph, enabling scalable offline evaluation with heterogeneous user models.
EDATracer provides an agentic framework for large-scale EDA artifact analysis. It targets cross-artifact debugging and design-flow understanding, offering an agent-based approach and benchmarking resources for evaluating EDA artifacts with LLM-enabled assistants.
D2F-ReAG tackles multi-hop reasoning augmented generation with dynamic decomposition and filtering. It improves retrieval-augmented generation by introducing dynamic problem decomposition and filtering to enable better cross-document reasoning and efficiency.
EASy proposes an efficient LLM-based agentic system. It focuses on optimizing execution efficiency under practical constraints (executor capability, computation cost) by improving context handling, task routing, dependencies, and feedback integration.
MediRec presents an explainable framework for Chinese medication recommendation using LLMs. It emphasizes interpretable clinical reasoning and justification to support medication decisions in Chinese-language health data, bridging explainability with clinical usefulness.
PICopilot proposes an LLM-based agent framework to assist photonic integrated circuit design through script generation. It aims to reduce the need for deep tool API proficiency and improve design productivity by producing executable design scripts.