Table 2 General advantages and disadvantages of graph neural networks.
From: Graph machine learning for integrated multi-omics analysis
Advantages |
State-of-the-art performance; versatility for a wide range of data types; can capture complex relationships and dependencies; well-suited when data context and connectivity are important; enhanced capability for transfer learning; scale efficiency for relatively large datasets; learning that factors in local and global information; can manage both graph structure and node/edge features; can combine information from proximal nodes/edges as well as the global context; extensions can handle heterogeneous graphs with different types of nodes and edges. |
Disadvantages |
Mainly can engage with static graphs; limited resilience to graph perturbations with incomplete or noisy data; computationally expensive on large graphs requiring significant resources for training; large amount of annotated data are required for training; model may lose information about specific nodes due to over-smoothing; can be challenging to generalize well on large graphs; challenging to find balance between incorporating prior information and learning from data; limited explainability unless extended to produce interpretable representations. |
