On the Power of Graph Neural Networks and Feature Augmentation Strategies to Classify Social Networks (2401.06048v1)

Abstract: This paper studies four Graph Neural Network architectures (GNNs) for a graph classification task on a synthetic dataset created using classic generative models of Network Science. Since the synthetic networks do not contain (node or edge) features, five different augmentation strategies (artificial feature types) are applied to nodes. All combinations of the 4 GNNs (GCN with Hierarchical and Global aggregation, GIN and GATv2) and the 5 feature types (constant 1, noise, degree, normalized degree and ID -- a vector of the number of cycles of various lengths) are studied and their performances compared as a function of the hidden dimension of artificial neural networks used in the GNNs. The generalisation ability of these models is also analysed using a second synthetic network dataset (containing networks of different sizes).Our results point towards the balanced importance of the computational power of the GNN architecture and the the information level provided by the artificial features. GNN architectures with higher computational power, like GIN and GATv2, perform well for most augmentation strategies. On the other hand, artificial features with higher information content, like ID or degree, not only consistently outperform other augmentation strategies, but can also help GNN architectures with lower computational power to achieve good performance.

• The paper demonstrates that integrating artificial features like node degree and identity vectors significantly improves classification performance across various GNN models.
• It compares four distinct GNN architectures, with advanced models such as GIN and GATv2 showing robust results even without inherent node data.
• The study highlights a trade-off between computational demand and feature richness, offering guidance for optimizing GNNs in practical network analysis.

Overview of Graph Neural Networks in Social Network Classification

Graph Neural Networks (GNNs) are instrumental in analyzing complex network structures, specifically in dealing with graph classification tasks. The research focuses on evaluating four different GNN architectures—GCN with Hierarchical and Global aggregation, GIN, and GATv2—across synthetic datasets designed to mimic classic network models without innate feature information. To compensate for the lack of inherent features within the nodes or edges of these networks, researchers have innovated by incorporating five different artificial feature augmentation strategies.

Augmentation Strategies and Their Impact

The artificial feature types developed for this paper include simple constants, random noise generation, node degree, normalized node degree, and a more complex identity vector that encapsulates cycles of various lengths associated with a node. These artificial features serve as a substitute for the actual node features that are not present in synthetic data. By experimenting with a combination of these feature types and GNN models, this research aims to probe the depths to which a GNN's computational capabilities can accurately classify graphs based on structure alone.

Comparative Performance Analysis

The paper illuminates the symbiotic relationship between the computational strength of a GNN and the level of information provided by the input features. Advanced GNNs like GIN and GATv2 showcase commendable performance across almost all feature types without much differentiation. However, the ID and degree features exhibit remarkable consistency in enhancing performance across all GNNs, further underscoring the value of richer information content. Importantly, while these more informative features inherently require more computation, the paper suggests that even less powerful GNNs can leverage them to attain respectable classification results.

Implications and Future Directions

In conclusion, the research signifies that the predictability of a GNN model on a classification task is not solely dependent on one factor, be it the neural network's architecture or the feature's informational depth. Both play integral roles in defining success. The findings are prelude to a larger discussion on the balance between computational demand and feature selection, which is pivotal in optimizing GNNs for real-world application. Future explorations are directed towards perfecting the hyperparameter optimization and applying these synthetic model-trained GNNs on real-world datasets for network classification, thereby extending the practical relevance of this paper to real social network analysis.