AdaGNN: Adaptive-Boosting Graph Neural Networks for Node Classification

Graph neural networks (GNNs) have become the de facto standard for learning over graph-structured data, yet they often falter when labels are scarce, graphs are noisy, or structural patterns are complex. While boosting is a powerful ensemble paradigm that sequentially corrects model errors, existing adaptations to graphs either break end-to-end training by introducing external boosters or resort to explicit sample reweighting that undermines relational inductive biases. To overcome these challenges, we introduce adaptive-boosting graph neural networks (AdaGNN), a fully integrated boosting framework designed specifically for graph learning. AdaGNN trains a series of GNN learners in a unified pipeline: each learner is warm-started from the previous one and ingests the progressively refined node embeddings. After each round, correctly classified nodes are frozen, i.e., their embeddings locked and excluded from subsequent updates, so that remaining learners focus solely on residual errors. This Embedding Freeze mechanism implicitly propagates error signals through the graph topology without requiring explicit weight adjustments or multi-stage coordination. Comprehensive evaluations across varied supervision levels demonstrate that AdaGNN consistently elevates classification accuracy and robustness, delivering substantial gains under limited labels while preserving or enhancing performance when labels are plentiful. AdaGNN establishes a novel, graph-native adaptive boosting paradigm—melding sequential GNN refinement, implicit error handling, and end-to-end training—to significantly advance the accuracy and robustness of GNN applications.