Table 7 Comparison of the GNN model with Advantages, Disadvantages, and application areas: [30]
Model | Features/characteristics | Message Passing Mechanism | Attention Mechanism | Aggregation Strategy | Scalability | Use Cases | Advantages | Disadvantages |
|---|---|---|---|---|---|---|---|---|
Graph Convolution Neural Network (GCN) | • Propagates information through neighbors • The simple and basic method • It uses a linear transformation • Homophily (focuses on immediate neighbors) • Smoothness assumption (neighbors should have similar representations) | Fixed weighted sum | No attention | Aggregation shown in | Limited | • Node classification • Semi-supervised Learning • Recommendation systems | • Simplicity & interpretability • Stable training often requires fewer epochs | • We have limited expressive power • Inability to capture local structures • No edge features |
Graph Attention Network (GAT) | • Learns weights for each neighbor's message • Used for transductive and inductive Learning, i.e., you can work with graph structures you've never seen before • Can handle varying neighborhood sizes | Weighted sum with learned weights | Self-attention | Aggregation | Moderate | • Node classification • Link prediction • Any task requiring localized information | • Ability to capture fine-grained relationships • Improved performance on tasks requiring attention to specific neighbors | • Computationally more expensive than GCN • It can be more sensitive to hyperparameters |
GraphSAGE | • Aggregates information from sampled neighbors • The number of model parameters is independent of the number of graph nodes. This makes GraphSAGE able to handle larger graphs • It can handle both supervised and unsupervised tasks • Sample-based approach with random or predefined sampling strategies | Fixed weighted sum | No attention | Aggregation | Highly scalable | • Node classification • Link prediction • Tasks where scalability is crucial | • Scalability to large graphs • Flexible sampling strategies • Suitable for graphs with varying node degrees | • Limited ability to capture global graph structures • May require more epochs to train effectively |