Time-aware Graph Attention Networks (TAGAT) for Multiperiod Default Prediction

This repository contains the implementation of the Time-aware Graph Attention Networks (TAGAT) for multiperiod default prediction in credit risk management, as presented in our paper "Time-aware Graph Attention Networks for Multiperiod Default Prediction".

Model architecture

Usage

Running the Scripts

quickly start running our TAGAT model, you can use the provided shell script run_GAT.sh.

./run_GAT.sh <device> <fold>

Replace <device> with the device identifier (e.g., cpu or the index of your cuda device) and <fold> with the specific fold number within the data folder (e.g., 1, 2, etc.).

Python Files Explanation

The shell script run_GAT.sh utilizes several Python files to train, validate, and evaluate the TAGAT model. Here's an explanation of the purpose of each Python file:

main_valid.py

This script is responsible for training and validating the model. It performs hyperparameter tuning and early stopping based on validation performance.

• Purpose: Find the optimal number of epochs based on validation data.
• Key Arguments:
  • --model_nam: Name of the model (e.g., NeuDP_GAT).
  • --device: Device to run the model on (e.g., cpu or the index of your cuda device).
  • --experiment_type: Type of how the data is preprocessed (e.g., expand_len).
  • --fold: Fold number of the data.
  • --max_epoch: Maximum number of epochs for training.
  • --patience: Patience for early stopping.
  • --batch_size: Batch size for training. In our case, it indicates how many timestamp of data are utilized during the training.
  • --learning_rate: Learning rate for the optimizer.
  • --weight_decay: Weight decay (L2 regularization).
  • --gamma: Exponential learning rate scheduler.
  • --lstm_num_units: Number of units in the RNN layer.
  • --cluster_setting: Clustering setting (e.g., industry).
  • --n_cluster: Number of clusters, which is the number of sectors in our case.
  • --intra_gat_hidn_dim: Hidden dimension for intra-sector GAT.
  • --inter_gat_hidn_dim: Hidden dimension for inter-sector GAT.

main.py

This script is responsible for training the model using the number of epochs determined during the validation phase.

• Purpose: Train the model using the full training set and the optimal number of epochs.
• Key Arguments: Same as main_valid.py, with the addition of:
  • --num_epochs: Number of epochs to train the model.

predict.py

This script is used for generating predictions on the test dataset using the trained model.

• Purpose: Generate predictions for the test dataset.
• Key Arguments: Same as main.py, with the addition of:
  • --data_file: The file containing the test data.
  • --restore_dir: Directory from which to restore the trained model weights.

report.py

This script generates a report comparing the predicted values with the actual values to evaluate model performance.

• Key Arguments: Same as main.py, with the addition of:
  • --pred_file: File containing the model's predictions.
  • --label_file: File containing the ground truth labels.

Generating Evaluation Metrics

The following commands in the shell script extract the Accuracy Ratio (AR) and Root Mean Square Normalized Error (RMSNE) metrics from the results:

python -c "import json; print('\n'.join(str(v) for k, v in json.load(open('$model_dir/test_metrics_pred_best_weights.json')).items() if k.startswith('cap')))" >> $model_dir/AR &&
python -c "import json; print('\n'.join(str(v) for k, v in json.load(open('$model_dir/test_metrics_pred_best_weights.json')).items() if k.startswith('recall')))" >> $model_dir/RMSNE  

• Purpose: Extract and save the AR and RMSNE metrics from the JSON file containing the evaluation results.