Showing 13 papers for 2026-07-27
We propose a traffic-signal control interface that uses a shared graph neural network to assign scores to individual traffic movements. Each junction then converts these scores into its own variable-sized set of legal signal phases via a deterministic incidence matrix. The design uses directed corridor nodes to provide context and movement nodes to represent input–output paths; learning focuses on movement scores while phase definitions and timing remain non-learned, enabling scalable control on heterogeneous road networks.
Coarsening-based training scales GNNs but hurts performance on heterophilic graphs. We show empirically and theoretically that existing coarsening methods lose crucial heterophily signals and degrade accuracy. We propose Adaptive Complementary Enhancement that preserves or recovers heterophily information during coarsening, improving performance on challenging graphs.
LatentFlow introduces visual analytics for exploring latent spaces of molecular GNNs. It helps chemists understand how molecular embeddings organize chemical information and reflect meaningful relationships, enabling diagnosis of model behavior beyond predictive accuracy. The tool supports interactive exploration of latent structure to reveal chemistries encoded by the model.
We address scalable graph-based recommendations with a combination of graph coarsening and label propagation. This approach constructs smaller, structurally representative graphs from large-scale systems to reduce computation while maintaining predictive quality. The method is flexible and applicable to industrial-scale recommendation problems.
We introduce Entropic Curvature, a global transport-based notion of curvature on graphs. Extending the Lott–Sturm–Villani framework to discrete graphs via the entropy along Wasserstein geodesics, it provides insights into how information propagates and helps address oversmoothing and oversquashing in GNNs.
GNNs rely on local message passing, which can hinder long-range propagation on graphs. We propose Spectral Flow Certificates (SFCs), inexpensive, single-scalar certificates computed from normalized Laplacians that reveal before training whether a graph supports depth-aware long-range information flow, without modifying or training models.
LayerNorm in GNNs often erases topology signals (degree, centrality). We analyze where in residual blocks this erasure happens and find that a positive per-node scalar inserted before LayerNorm is canceled by normalization, while the same scalar after LayerNorm affects the representation magnitude; the surviving slot is the post-LayerNorm position. PostDeg provides a parameter-free post-LayerNorm placement to preserve topology signals.
We present a general GNN framework for learning algebraic properties from Cayley graphs of finite groups. The pipeline unifies graph construction, features, training, and architecture, while different tasks only change the target labeling function. We demonstrate learning properties such as abelianity and nilpotence across diverse Cayley graphs.
We combine graph-based molecular fragmentation with covariant direct learning of forces to enable coupled-cluster accuracy AIMD for fluxional systems. The fragmentation reduces computational cost and avoids reliance on automatic differentiation on energy surfaces, while preserving accuracy with direct force learning. This enables CC-quality dynamics for challenging reactive systems.
MineValiCoder tackles reliable code generation with test-case quality mining and bipartite graph-based mutual validation. This reduces dependence on hand-crafted tests and mitigates stochasticity in LLM outputs, providing a more robust evaluation signal for generated code.
MedKGent is an LLM-agent framework that builds temporally evolving medical knowledge graphs. Using over 10 million PubMed abstracts from 1975 to 2023, MedKGent incrementally constructs a KG daily via two specialized agents. The Extractor Agent identifies entities and relations; the Builder Agent integrates them over time.
RankGraph-2 presents lifecycle co-design for billion-node graph learning in recommendation. Meta-scale deployment jointly optimizes graph construction, representation learning, and real-time serving. A co-trained cluster index enables scalable online retrieval without expensive online KNN, showing how to integrate the three stages.
Ghost Vectors investigates the privacy implications of soft-deleted embeddings in HNSW vector databases used by RAG. Embeddings can remain reconstructible after a deletion request, raising GDPR/HIPAA concerns. The paper analyzes reconstruction risks and discusses mitigation strategies.