Graph Neural Networks for Graph Classification on OGBG-PPA

This repository contains a collection of graph neural network (GNN) models implemented and evaluated for graph classification tasks using the OGBG-PPA dataset. The work explores and compares various architectures such as GCN and GIN, applying different pooling and aggregation strategies.

📦 Models Implemented

All models are built using PyTorch Geometric. Each model can be configured with a wide variety of training parameters.

• Baseline (Optional model without graph neural network architecture to compare )
• Graph Convolutional Network (GCN)
• Graph Isomorphism Network (GIN)
• Custom configurations:
  • Number of layers
  • Node representation dimensionality
  • Aggregation method (sum, mean, max, add)
  • Dropout rate
  • Use of batch normalization
  • Use of residual connections
  • Pooling method (mean, add, max, top-k, etc.)

🔧 Training Techniques

• Batch Normalization: optional, improves training stability
• Residual Connections: optional, helps deeper networks
• Dropout: configurable dropout ratio

📊 Best Results (Summary)

The models were evaluated using the standarized evaluator from the package.

Model	Layers	Dim	Pooling	Batch Norm	Residual	LR	Dropout	Epsilon	Val Acc.	Test Acc.	Epochs	Time (GPU)
Baseline	3	128	mean	❌	❌	0.001	0.2	-	~11.9%	~13.3%	25	30 min
GCN	5	300	mean	✅	❌	0.001	0.2	-	~61.5%	~65.3%	60	4h 44min
GIN	4	200	mean	✅	❌	0.001	0.2	0.01	~68.2%	~73.1%	60	3h 24min

The reported training times were obtained using:

• GPU: NVIDIA GeForce RTX 3080 (with CUDA)
• CPU: Intel® Core™ i7-10700K @ 3.80GHz
• RAM: 32 GB DDR4 3200 MHz

📈 Test Accuracy (Mean ± Standard Deviation)

To account for variability between runs, the following results are reported as the mean ± standard deviation of test accuracy over multiple executions:

Model	Test Accuracy (Mean ± Std)
GCN	64.29 ± 0.002 %
GIN	72.23 ± 0.007 %

⚙️ Installation

pip install -r requirements.txt

For showing graphs

pip install matplotlib