Random Forest - Fraud Detection

This project uses a collection of Decision Trees in a Random Forest configuration to detect fraudulent charges. It was created as a project for the UNM CS429 - Intro to Machine Learning course in Spring 2024.

Developers

• Michael Millar: Sole developer

How to use

Setup

1. Have python 3 and pip installed on your computer.
2. With pip installed run the command pip install -r requirements.txt to install dependencies.

Config

There are several configuration options you can specify before running the application. They can all be changed from fraud.py file.

• mode: either 'PROD' or 'TEST'. Test just runs on the training data and does not create an output file for kaggle submission.
• model: either 'TREE' or 'FOREST'. Tree will run a single tree regardless of number of trees specified, while forest will train and test multiple trees as specified.
• sample_size: If you want to reduce the total samples of the original data set to a set smaller number. Leave set to None otherwise.
• save: If the model should be saved to a file. Saved models with numerical data may have issues being reloaded (infinite size).
• undersample: either True or False. If true, will resample the training data based on the proportions specified in undersample_props.
• undersample_props: An array-like value which indicates the proportions of each target class. Should sum to 1.
• total_trees: Total number of trees to create in the random forest.
• total_depth: The maximum depth any tree can grow to.
• sample_cutoff: The minimum samples required to branch.
• confidence_alpha: Confidence level between 0 and 1.
• gain_func: Information gain metric to use, valid options are impurity.entropy, impurity.giniIndex, and impurity.misclassificationIndex.
• train_file: Path to the training data
• test_file: Path to the testing data

Running the Program

To run the program you just need to have python run the fraud.py file. This can be done with the command: python fraud.py. You may need to use python3 depending on your setup.

---

Kaggle Competition Result

In the Kaggle competition, this algorithm had the highest balanced accuracy score of 77.06 on 3/9/24. This was achieved with a single decision tree using entropy as information gain metric, min samples of 10, max depth of 25, and a confidence value of 0.95.

The results from the random forest always produced lower accuracy ratings, despite the individual trees of the random forest having higher accuracy.

---

File Manifest

• .gitignore: Hide all the unwanted files from the git repository.
• README.md: This description file.
• classifier.py: This contains the logic of the DecisionTreeClassifier and the Node class. This is the class that builds, trains, and predicts with.
• fraud.py: The starting point of the project. This file loads and prepares the data, initializes the classifiers, and starts the train and test processes.
• impurity.py: This file contains the entropy, gini index, misclassification index, information gain, and chi-square methods.
• random_forest.py: This file contains the RandomForestClassifier and creates the decision trees in the random forests.
• requirements.txt: The file that contains all the additional libraries required to run the script.
• utils.py: This file contains some utility helper functions that get used multiple times.