Showing 20 papers for 2026-07-30
Permutation equivariance in graph neural networks makes attribution scores invariant under input automorphisms, so two chemically equivalent nitro groups receive identical attributions. Consequently, top-k explanations are not canonical and one edge is arbitrarily named due to array ordering; this is a structural obstruction rather than a bug, calling for a canonical explanation among automorphic alternatives.
This work investigates the effectiveness of message-passing GNNs for regression tasks. While GNNs excel in graph-structured data like molecules and networks, benchmarking has largely focused on classification; the paper assesses regression performance, identifies evaluation gaps, and proposes appropriate protocols and baselines for regression settings.
We survey neural architecture search (NAS) methods for traffic prediction, outlining search spaces, objectives, and performance. The review highlights challenges such as spatio-temporal heterogeneity and cross-city generalization, and sketches future directions for NAS-driven traffic forecasting.
AgentGFM introduces a Graph Foundation Model with node-agent information-flow control, enabling adaptive propagation strategies rather than fixed schemes. By learning how information travels conditioned on graph context, the model improves transferability across domains and robustness to local structural variation.
The paper proves universality results for permutation-equivariant neural networks with partially random node features. It shows that PENNs with some random features can approximate arbitrary measurable functions in probability, providing rates and guidance on how random features enhance expressivity for GNNs.
This work presents visibility-aware graph-based inference for business conduct risk under sparse and biased data. By leveraging inter-firm networks to propagate risk signals, the model improves predictions of future recorded conduct events and mitigates the impact of incomplete incident data.
ReDiSC introduces a reparameterized masked diffusion model for scalable node classification that accounts for structured label dependencies. By modeling joint label distributions within a diffusion framework, it enables scalable, accurate predictions on large graphs.
The paper analyzes the Rademacher complexity of graph neural networks to unify expressivity and geometry, revealing a trade-off between expressive power and generalization. It connects WL-based expressivity with learning guarantees, providing theoretical bounds for GNNs.
MapTab is a diagnostic benchmark for evaluating multimodal reasoning in language models on long-horizon route-planning tasks that require grounding map visuals and route attributes. It assesses holistic multi-criteria reasoning in heterogeneous topologies.
GEqTrain is a configuration-driven framework that retargets equivariant GNNs across 3D scientific tasks by decoupling dataset semantics, model composition, and training objectives. It maps raw data to typed fields and assembles models and losses via declarative configurations to improve reusability.
PINCO combines physics-informed neural networks with graph neural networks to solve AC-OPF robustly on unfiltered data, including ill-conditioned cases. It addresses topology changes up to N-2 contingencies and can detect feasible OPF instances, offering unsupervised robustness.
Graph Signal Diffusion Models for wireless resource allocation train a diffusion policy to match expert conditional distributions over allocations. Using a primal-dual approach, it generates primal iterates as samples from these conditionals, treating allocations as stochastic graph signals on channel state graphs.
HYVINT presents intensity-driven hypergraph generation with variational embeddings, offering an interpretable mechanism to model polyadic incidences in hypergraphs and improving realism and controllability of generated structures.
This study compares five image encoders within GCN frameworks for breast ultrasound classification, examining how encoder choice affects graph construction, homophily assumptions, and downstream accuracy. It offers practical guidance on encoder-graph design in medical imaging.
UrbanDS introduces a graph-guided multi-agent system that coordinates LLMs for data-intensive urban tasks, enabling discovery, integration, and reasoning over large, heterogeneous urban datasets. The approach improves data utilization and task performance in complex urban analytics.
This methodology provides a structured process for designing knowledge-driven missions in ROS 2, including defining initial and target conditions, task structuring, sequencing, knowledge graph data representation, and high-level mission design language.
WhisperRec offers a latent reasoning framework for foundation recommendation models that replaces explicit chain-of-thought with implicit latent reasoning. This reduces inference overhead while preserving recommendation quality.
The paper develops methods to detect inconsistencies across text, tables, and knowledge graphs, presenting a taxonomy of cross-modal conflicts and proposing practical approaches to detect and explain such disagreements.
Gwhere is a generative next-POI recommender for Amap that generates POI identifiers to predict the next destination, integrating spatial structure and heterogeneous signals for scalable, coherent recommendations.
IMFuse presents an instance-aware multi-layer fusion approach for LLM-enhanced sequential recommendation, leveraging semantic signals across multiple LLM layers to overcome dimensional collapse in final-layer representations.