Graph Convolutional Network on data from Elliptic bitcoin dataset of transactions graph

• Python
• PyTorch
• pandas
• scikit-learn

• Download the Elliptic Dataset for Bitcoin transactions
• Don't change the names of the csv files
• If you change the name of the folder pass the path with the changed name as the command line argument for the dataset directory while training and testing

The files present in Code are:

• util.py: Contains the function for loading the data
• model.py: Implementation of the Graph Convolutional Network with 2 layers
• train_GCN.py: Python script to train the GCN
• test_GCN.py: Python script to test the GCN using weights from training
• train_SkipGCN.py: Python script to train the SkipGCN
• test_SkipGCN.py: Python script to test the SkipGCN using weights from training
• GCN_Elliptic Dataset.ipynb: The main ipynb notebook used for all the tasks

The weights for the models are provided in a folder titled gcn_weights. They can be obtained here

python train_GCN.py -d [:dataset directory path] -e [:number of epochs] -l [:learning rate] -t [:number of timesteps to train] -m [:directory to save model weights]

python test_GCN.py -d [:dataset directory path] -t [:timestep to start testing] -m [:model weights directory]

python train_SkipGCN.py -d [:dataset directory path] -e [:number of epochs] -l [:learning rate] -hs [:hidden layer size] -t [:number of timesteps to train] -m [:directory to save model weights]

Training the SkipGCN might be unstable and lead to NANs in the output. The problem might be solved by changing the size of the hidden layer. Use the -hs argument to change the hidden layer size. Use the same hidden layer size for training and testing. The model weights provided are for a hidden layer of size 16.

python test_SkipGCN.py -d [:dataset directory path] -t [:timestep to start testing] -hs [:hidden layer size] -m [:model weights directory]