Evaluating Deep Graph Network Performance by Augmenting Node Features with Structural Features

Mohamad Elhadi Abushofa, Amir Atapour-Abarghouei, Matthew Forshaw, A. Stephen McGough

Abstract: In machine learning, features play a vital role in modeling and understanding the data; their quality and representation essentially determine how accurate the results are. The problem is compounded in the graph-based learning paradigm when one considers how complex and interconnected the data is. However, to achieve more accurate results, augmenting graph data poses specific challenges in the field of graph learning. Feature augmentation is a critical aspect of enhancing data. Moreover, some datasets have limited features and some real datasets do not have features. In this paper, we present our approach termed Feat-Aug which is an extension of our previous work on non-parametric approaches. The aim of this work is to augment node features in graphs on parametric approaches such as Graph Neural Networks (GNNs) with the objective of improving performance in node classification tasks. Our approach combines real features, such as a bag of words in citation networks, which are typically associated with nodes, with structural features extracted at the node level, such as node degree and clustering coefficient. To further enhance these features, we leverage deep learning models to incorporate additional node-level features. The final modified features are the result of the combination of both real and structural features. To evaluate the effectiveness of the approach, we carried out extensive experiments with several real datasets. Moreover, our method consistently outperforms or achieves comparable results to Graph Neural Networks (GNNs) baselines and their variations, such as popular graph neural network models. Crucially, our approach deals with the problem of insufficient real features in certain datasets. This study is a major progression in the field through an effective node classification model. By integrating both real and structural features, our approach holds promise to improve the performance of node classification models.