weekly-customs-selection

Apply DATE model for Weekly Customs Selection (or Daily)

Written by: Sundong Kim, Institute for Basic Science

Preparation

• First, locate the training data under the directory ./data/
• Second, locate the weekly dataset under the directory ./weekly-data/, the weekly dataset name should follow week*_ano.csv

Prediction

• To make a prediction, run pilot.sh ($ bash pilot.sh). By running this bash script, XGBoost classifier, XGBoost + Logistic regression, and two variants of our DATE model predict the test data.
• The prediction result will be saved under the './weekly-data/' with the file name of 'week*_ano_result'.

Parameter tuning

• You can adjust some parameters in the script, you can refer the arguments of preprocess_data.py, generate_loader.py, and train.py, which are listed just below the main part. (e.g., You can change the learning rate of the DATE model by adding --lr 0.01)
• In our setting, we consider training period from 2016-01-01 ~ 2017-10-01 and 2017-10-01 ~ 2017-12-31 as a validation period.

Results

Considering the total number of imports are over two thousands, Top-10 high fraudulent transactions, and bottom-10 low fraudulent transactions selected by DATE and other models are quite consistent each other. Check this Jupyter Notebook for more details.