Rini Widiastuti, Aji Prasetya Wibawa, Wahyu Sakti Gunawan Irianto, Jehad A. H. Hammad
Graph Convolutional Networks (GCNs) are widely used for node classification because they combine node features and graph topology effectively. However, their performance can be limited by structural noise, over smoothing, and sensitivity to graph characteristics. Ensemble learning is known to improve predictive performance, but its independent contribution to GCN performance remains unclear. This study presents a systematic empirical evaluation of five ensemble strategies (bagging, boosting, voting, averaging, and stacking) applied to GCNs under identical base architectures and controlled settings. Experiments were conducted on ten benchmark graph datasets with diverse structural properties. Performance was evaluated using accuracy, precision, recall, and F1-score. The results show that ensemble methods can improve GCN performance, especially on datasets with larger scale and more complex structure. Their benefits are smaller on simpler graphs. Among the evaluated methods, stacking provides the most consistent overall performance. Bagging, voting, and averaging also show stable and competitive results. Boosting remains more sensitive to graph characteristics and shows less consistent behavior than the other strategies. These findings clarify when ensemble learning is beneficial for GCNs and provide guidance for developing more robust graph neural network systems. © 2013 IEEE.
Universitas Negeri Malang, Faculty of Engineering, Department of Electrical and Informatics Engineering, Malang, 65145, Indonesia; Al-Quds Open University (QOU), Faculty of Technology and Applied Sciences, West Bank, Ramallah, Palestine
