Determining the optimal number of GAT and GCN layers for node classification in graph neural networks
Node classification in complex networks plays an important role including social network analysis and recommendation systems. Some graph neural networks such as Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) have emerged as effective approaches for achieving high-performance c...
| Main Authors: | , , , , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English |
| Published: |
Institute of Electrical and Electronics Engineers Inc.
2023
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/40357/1/Determining%20the%20optimal%20number%20of%20GAT%20and%20GCN.pdf http://umpir.ump.edu.my/id/eprint/40357/2/Determining%20the%20optimal%20number%20of%20GAT%20and%20GCN%20layers%20for%20node%20classification%20in%20graph%20neural%20networks_ABS.pdf |
| Summary: | Node classification in complex networks plays an important role including social network analysis and recommendation systems. Some graph neural networks such as Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) have emerged as effective approaches for achieving high-performance classification in such tasks. However, constructing a graph neural network architecture is challenging particularly due to the complex task of determining the optimal number of layers. This study presents a mathematical formula for determining the optimal number of GCN and GAT hidden layers. The experiment was conducted on ten benchmark datasets, evaluating performance metrices such as accuracy, precision, recall, F1-score, and MCC for identifying the best estimation of number of hidden layers. According to the experimental findings, the number of GAT and GCN layers selected has a substantial impact on classification accuracy. Studies show that adding extra layers after the optimum number of layers has a negative or no impact on the classification performance. Our proposed approximation technique may provide valuable insights for enhancing efficiency and accuracy of the Graph Neural Network algorithms. |
|---|