Fraud Detection

The objective of this machine learning model is detect fraudulent transactions in the financial services sector.

Stakeholders

People involved in this project

Role	Responsibility	Full name	e-mail
Data Scientist	Scrum Master	Fernando Felix	fernandofelix@copin.ufcg.edu.br
Data Scientist	Author	Dunfrey P. Aragão	dunfrey@gmail.com
Data Scientist	Author	Ítalo de Pontes Oliveira	italodepontesoliveira93@gmail.com

Usage

Clone the repository:

git clone https://github.com/<@github_username>/fraud_detection.git
cd fraud_detection

Prerequisites

We have tested the library in Ubuntu 19.04, 18.04 and 16.04, but it should be easy to compile in other platforms.

Python

The project was written in Python 3, and work with later as well. Also, please read up the subsequent libraries that are used: CatBoost, Scikit Learn, PyOD, Pandas and PySpark.

Libraries

Please make sure that it has installed all the required dependencies. A list of items to be installed using pip install can be running as following:

pip install -r requirements.txt

All Environment

It is possible to use the file install.sh to install all requirements automatically.

Execute the script on the shell:

$ chmod +x install.sh
$ sudo ./install.sh

Running

The project does not need user interaction of information in time execution. To run the whole workflow of the project, it is possible by the following command:

$ python main.py run \
  --input_train_file ../data/xente_fraud_detection_train.csv \
  --input_test_file ../data/xente_fraud_detection_test.csv \
  --output_balanced_train_x_file ../data/balanced_train_x.csv \
  --output_balanced_train_y_file ../data/balanced_train_y.csv \
  --output_valid_x_file ../data/valid_x.csv \
  --output_valid_y_file ../data/valid_y.csv \
  --output_valid_result_file ../data/valid_result.csv \
  --output_test_result_file ../data/xente_output_final.txt

• input_train_file & input_test_file: train and test datasets available by Zindi.
• output_balanced_train_x_file & output_balanced_train_y_file: 70% of the oversampled training dataset to work on the training step.
• output_valid_x_file & output_valid_y_file: 30% of the training dataset saved on memory to be used in validation step.
• output_valid_result_file: outcome from the classification model using validation data.
• output_test_result_file: outcome from the classification model using test data.

If that interested in work just a specific step (train, validate, or test), it is possible by:

python <mode> <--key> <keywords>

• mode: run, train, validate, and test
• --key: input_train_file, input_test_file, output_balanced_train_x_file, output_balanced_train_y_file, output_valid_x_file, output_valid_y_file, output_valid_result_file, output_test_result_file
• keywords: file name in your choice

For example:

$ python main.py train \
  --input_train_file ../data/xente_fraud_detection_train.csv \
  --output_balanced_train_x_file ../data/balanced_train_x.csv \
  --output_balanced_train_y_file ../data/balanced_train_y.csv \
  --output_valid_x_file ../data/valid_x.csv \
  --output_valid_y_file ../data/valid_y.csv

The following scripts want to have the files to use by the model and the model created to be successful. In the test, it is necessary to have the outlier detectors trained. All of them can be made at the training step.

$ python main.py validate \
  --output_valid_x_file ../data/valid_x.csv \
  --output_valid_y_file ../data/valid_y.csv \
  --output_valid_result_file ../data/valid_result.csv

$ python main.py test \
  --input_test_file ../data/xente_fraud_detection_test.csv \
  --output_test_result_file ../data/xente_output_final.txt

Model Frequency

The entire process is running over 11'44". To training and predict the outcomes by detector outliers, it happens over 2'42," and to train dataset oversampling happens over by 4'10".

Evaluation

On-line: the project is evaluated on top 70, over +2000 submissions.

The project has been evaluated by a Zindi online application, according to the F1 score value, which ranges from 0 (total failure) to 1 (perfect score). Hence, the closer score is to 1, the better is the model. This project reviewed before the submission and respecting all criteria "meet specifications" in order to pass.

Off-line: The project submitted off-line to validate success using the dataset has saved 44.2% of lost by fraud/genuine transaction.

Documentation

• project_specification.md: gives a data-science oriented description of the project.

• Fraud Detection Presentation and Classification Model Solution

• Overview Task and System Documentation (Written in Portuguese)

Folder structure

Project Folder Structure and Files

• data: contains data files generated and used by the classification model.
• docs: contains documentation of the project
• fraud_detection: main Python package with source of the model.
• jupyter-notebook: contains jupyter notebooks evaluation and modeling experimentation.