GraphFC: Customs Fraud Detection with Label Scarcity

This repo contains the PyTorch implementation for "GraphFC: Customs Fraud Detection with Label Scarcity".
The paper along with performance analysis on three real customs datasets can found here

Model Architecture of GraphFC

Model architecture of GraphFC. Cross features extracted from GBDT step act as node features in the transaction graph. In the pre-training stage, GraphFC learns the model weights and refine the transaction representations. Afterwards, the model is fine-tuned with labeled data with dual-task learning framework to predict the illicitness and the additional revenue.

How to train the model

The model code for GraphFC lies in graph_sage directory. Simply run graph_sage/train.py and specify the dataset parameters could train the model and evaluate the performance. Please refer to the scripts under the directory run_*Data.sh for reproduce the results for individual country.

graph_sage
   |-- dataset.py -> Preprocess for customs data
   |-- models.py -> Main model modules
   |-- parser.py -> training arguments
   |-- pygData_util.py -> Data structure for graph data
   |-- run_Mdata.sh
   |-- run_Ndata.sh
   |-- run_Tdata.sh
   |-- train.py -> Train model
   |-- utils.py

Arguments and Hyperparameters

# Dataset parameters
--data: Country name for building dataset ['synthetic', 'real-n', 'real-m', 'real-t']
--initial_inspection_rate: Initial inspection rate of labeled data
--train_from: Starting date of training data
--test_from: Starting date of testing data
--test_length: Number of days for testing data


# GraphFC Hyperparameters
--seed: Random seed
--epoch: number of epochs
--l2: l2 regularization 
--dim: dimension for hidden layers 
--lr: learning rate
--device: The device name for training, if train with cpu, please use:"cpu" 

Data

You can experiment with GraphFC by downloading synthetic customs data from this repo.