Showing 13 papers for 2026-04-06
We propose a fairness-aware training method for graph neural networks by enhancing the counterfactual augmented fair GNN framework (CAF) with homophily awareness. The model uses supervised contrastive learning and counterfactual augmentation to reduce biases arising from node attributes and graph structure, improving fairness without sacrificing performance. The approach is evaluated on standard benchmarks and shows better fairness-accuracy tradeoffs than prior methods.
We introduce a complex-valued GNN parametrization that is globally invariant to local basis choice, enabling distributed control of planar networked systems in GPS- and compass-denied environments. The method handles 2D geometric features and basis transformations, ensuring consistent behavior across heterogeneous sensors. This yields robust, scalable, and basis-agnostic controllers for planar graphs.
Non-Exemplar Continual Graph Learning (NECGL) suffers feature drift because it stores only prototypes. We propose Analytic Drift Resister (ADR), which exploits the generalization properties of frozen pre-trained models to stabilize representations without exemplars. The method provides a privacy-preserving alternative to rehearsal while improving continual learning performance.
ScaleGNN introduces a 4D hybrid parallelism and communication-free sampling to scale mini-batch GNN training. It reduces sampling bottlenecks and enables effective data, model, and feature parallelism, achieving near-linear scaling on large graphs. The framework broadens the practicality of GNN training for industrial-scale datasets.
We develop a Graph Neural Network-based multimesh surrogate for fast flood forecasting. Built on a production Telemac2D model on a high-resolution mesh, the surrogate learns across multiple mesh scales to speed up predictions on operational domain. Applied to the lower Têt River (France), it delivers rapid forecasts with accuracy close to the high-fidelity solver.
DSBD proposes Dual-Aligned Structural Basis Distillation for Graph Domain Adaptation to address topology shifts. It emphasizes aligning structural bases—not only feature distributions—across source and target graphs to preserve geometric and spectral properties. This yields more reliable transfer and improved GDA performance under significant topology changes.
We propose an efficient LLM-based framework for causal graph discovery that uses a breadth-first search to achieve linear query complexity, improving over pairwise query strategies. The approach can incorporate observational data when available to further enhance accuracy. This makes causal discovery with LLMs scalable to larger graphs.
Graph Convolutional Branch and Bound introduces neural heuristics to guide B&B for NP-hard problems. The supplementary materials describe learning an optimality score with neural networks to evaluate nodes, enabling more efficient pruning. This integration accelerates search while improving solution quality.
ChronoSpike introduces an adaptive spiking GNN for dynamic graphs, combining event-driven spiking efficiency with global context. It addresses trade-offs of attention-based methods and recurrent models by using adaptive spiking to capture temporal evolution with reduced complexity. The approach improves scalability and performance on dynamic graphs.
OntoKG presents an ontology-oriented approach to constructing knowledge graphs, designing the schema early for ontology analysis with intrinsic-relational routing. By deciding entity types, edge semantics, and schema constraints upfront, it improves reuse and ontology-level reasoning downstream. The method aims to produce more coherent, analyzable KGs.
This workshop summary surveys progress in integrating large language models with graph data management and graph learning. It identifies key challenges, opportunities, and future directions for algorithms and systems that bridge LLMs and graph-structured data. It emphasizes cross-disciplinary collaboration and practical deployment considerations.
LogicPoison reveals logical attacks on Graph Retrieval-Augmented Generation, showing GraphRAG systems rely on graph topology for grounding and can be compromised by topology manipulation. The paper analyzes attack vectors grounded in topology and discusses potential defenses to preserve grounding integrity.
We propose User-Aware Conditional Generative Total Correlation Learning for multi-modal recommendation, arguing that content relevance to user preferences is not universal. The method models conditional total correlation between user interactions and multi-modal item content and tailors generation to individuals, reducing modality-specific noise and improving recommendation quality.