Showing 10 papers for 2026-06-26
Money Laundering Group Discovery (MLGD) aims to identify hidden criminal groups and recover their complete structures in large-scale financial networks. Existing graph anomaly detection methods mainly provide node-level risk alerts, while global group search remains passive. We propose a clue-guided framework that starts from concrete investigative clues and iteratively expands to recover the entire group structure, better matching AML workflows.
We generalize a previous graph neural approach and introduce a property-independent framework for learning algebraic properties of finite groups directly from Cayley graphs. The framework uses a common GNN architecture and training pipeline to study properties such as abelianity, nilpotency, and solvability. Through representative case studies, we assess how much algebraic structure can be recovered from graph representations.
We develop a quantitative theory for zero-shot size transfer for Neural ODEs on sparse random graphs by leveraging graphon limits. We define Graphon-NDEs and adjoint Graphon-NDEs as the infinite-node limits of forward and backward dynamics, enabling analysis of size transfer and stability. The results identify conditions under which a model trained on a small graph generalizes to larger, similar graphs.
We propose an explanation method for Event-based Temporal Graph Neural Networks (ETGNNs) based on feature-induced information flow. Unlike prior methods that inspect only a subset of the information paths, our approach traces the full flow of temporal features through the model to identify key pathways that drive predictions. The method improves explainability across diverse ETGNN applications.
This survey organizes Graph Neural Networks applications around a single design space and derives spectral and spatial formulations from first principles. It connects expressive power to the Weisfeiler-Leman hierarchy and clarifies when graph structure yields gains and when it does not. The result is a practical framework for understanding where GNNs are most effective.
We develop scalable message-passing quantum graph neural networks within the Weisfeiler-Leman framework. The work links quantum graph representations to classical GNNs and addresses scalability and trainability challenges of variational quantum circuits. These contributions enable more scalable and reliable quantum graph learning.
We study rotary position encodings (RoPE) for graphs and show that rotating token representations according to the graph Laplacian spectrum can efficiently inject structural information into attention. The resulting Wave-Induced Rotary Encodings (WIRE) improve performance on synthetic and real-world graph tasks, supported by theoretical insights about why the approach works.
We introduce MKG-RAG-Bench, a benchmark designed to evaluate retrieval in multimodal knowledge graph augmented generation. It targets the retrieval bottlenecks arising from heterogeneous multimodal knowledge and assesses cross-domain retrieval and grounding capabilities. The benchmark provides a gap-aware framework for evaluating retrievers in MKG-RAG settings.
EvoOptiGraph is a weakness-driven coevolution framework for optimization modeling. It treats each MILP as an attributed bipartite graph and uses validity-preserving evolutionary operators to generate structurally diverse problem instances. The approach couples data generation and model learning to continuously address model weaknesses.
TGHE proposes a template-based, ego-centric graph homomorphic encryption framework for privacy-preserving GNN inference in edge-cloud systems. By canonicalizing ego-graphs into a small set of structural templates at the edge and packing their representations, TGHE decouples per-query cost from full graph size and enables scalable encrypted computation on large, dynamic graphs.