Researchers introduced HarmonyGNN, a new training technique for graph neural networks (GNNs) that improves accuracy on heterogeneous graph data without relying on labeled nodes. Tests on 11 benchmark graphs showed state-of-the-art performance on seven homogeneous graphs and accuracy gains of 1.27% to 9.6% on four heterogeneous graphs. The framework also boosts computational efficiency. The work will be presented at the International Conference on Learning Representations in 2026.