SPArk Fraud Detection Model

Team Unbeatabolt ⚡️

Problem Statement

The institution wishes to find a solution to reduce fraudulent transactions while cutting down on false positives. Our team has developed the SPArk business plan, which comprises of Scheme tokenization, Predictive modelling, and AI fraud flagging. This repository addresses this second component, containing notebooks for data analysis and the predictive model:

• Descriptive Analysis (EDA): Informs the implementation of the business plan detailed in our slide deck, as well as feature and model selection.
• train_test: Notebook details model training, testing, feature selection, and hyperparameter tuning. A corresponding performance report is also attached.

Project Structure

The repository is organized as follows:

SPARKFraudModel/ ├── data/ # Datasets (raw and clean) used for training and eval ├── eda/ # Jupyter notebooks for exploratory data analysis ├── models/ # Saved models ├── slide-deck/ # Visualizations and pitch deck ├── README.md # Project README

Made as part of BOLT UBC's 2024 Case Competition. Load using Jupyter Notebook or Google Colab.

Data Description

The dataset used for this project contains anonymized financial transaction data, including features such as transaction amount, time, and various user-specific attributes. Due to privacy concerns, sensitive information has been excluded or anonymized.

Transaction ID: Unique identifier for each transaction. Amount: The monetary value of the transaction. Timestamp: The time when the transaction was made. Features: Various attributes that describe the transaction and the user. The data is stored in the data/ directory and is processed in batches using Spark DataFrames.

Feature Engineering

Feature engineering is a critical part of this project. Some of the key features created include:

• Transaction Frequency: Number of transactions made by a user within a specific timeframe.
• Monetary Ratio: Ratio of the current transaction amount to the user's average transaction amount.
• Time-based Features: Extracted hour, day of the week, and other temporal features to capture behavioral patterns.

Modeling

The model pipeline consists of the following steps:

1. Data Preprocessing: Handling missing values, normalization, and splitting data into training and testing sets.
2. Feature Selection: Using techniques like PCA to select the most relevant features. Insights taken from the EDA–such as which features were most correlated–was also implemented.
3. Model Training: Several models were trained, including Random Forest, Gradient Boosting, and a custom ensemble model.
4. Hyperparameter Tuning: Grid search and cross-validation were used to optimize the models.

Evaluation

The models were evaluated using various metrics, including:

• Accuracy: The proportion of correctly identified transactions.
• Precision: The key metric in reducing false alarms. The ratio of true positives to the sum of true and false positives.
• Recall: The ratio of true positives to the sum of true positives and false negatives.
• F1 Score: The harmonic mean of precision and recall, used to balance the two metrics.

Results

The final model achieved an F1 score of 0.92 on the test dataset, indicating high accuracy in detecting fraudulent transactions, and a high precision expected to minimize false alarms. The model's performance is visualized in the results/ directory, where confusion matrices and other evaluation plots are found.