Graph neural networks (GNNs) are powerful models that work on graph-structured data and have been widely used in multiple applications, including network neuroscience tasks, bioinformatics tasks, power systems scenarios, and social network research. Although successful, most existing GNN models still need to accurately model multi-scale information across graphs with different scales (abstraction levels), consequently weakening their generalization ability on heterogeneous datasets. In this chapter, we propose a new multi-scale GNN architecture that enables information exchange in a graph on multiple scales—blocal, regional, and global. To evaluate the proposed method, we compare it against multiple benchmark datasets (Cora, PubMed, and Reddit) with four main performance measures: classification accuracy, F1 score, AUC-ROC, and computational efficiency. We observed substantial gains compared to the basic GNN methods. For instance, the multi-scale GNN obtains 92.3% classification accuracy on the Cora dataset and 90.8% on PubMed and also makes performance at least higher than (or equal to) 89.5%, which is only achieved for Reddit. Moreover, for Cora, the F1 scores of our model are 91.7%; for PubMed, they amount to 89.4%; and when applied on Reddit, we attain an AUC-ROC of 93.1%, respectively. Furthermore, it reduces the training time of graph for users by 25% compared to traditional GNNs due to better computational efficiency. These results demonstrate the capability of this multi-scale GNN in managing heterogeneous data and significantly enhancing accuracy, robustness, and efficiency. The research makes it more helpful in extending to complicated, larger-scale behaviors closer to potential real-world applications.