Financial-Fraud-Detective

Artwork created by AI

Project Overview

This project aims to build a machine learning pipeline to detect fraudulent transactions in a synthetic dataset of bank transactions. The pipeline includes steps such as exploratory data analysis (EDA), data cleaning, model creation, hyperparameter tuning, and model evaluation.

Background

Financial fraud detection involves identifying fraudulent transactions within a dataset that predominantly consists of legitimate transactions. This project addresses the challenge of imbalanced datasets, where fraudulent transactions are the minority class. The goal is to create a model that effectively captures fraudulent transactions while minimizing false positives.

Project Structure

The project follows a structured approach with the following notebooks:

1. Initial EDA:

   • Conduct univariate, bivariate, and multivariate exploratory analysis.
   • Formulate hypotheses based on insights gained from the analysis.
   • Explore relationships between predictors and the target variable.

2. Data Cleaning and Pre-processing:

   • Clean and wrangle the dataset.
   • Handle missing values, outliers, and incorrectly formatted data.
   • Drop unnecessary columns based on EDA findings.
   • Save the pre-processed dataset for further analysis.

3. Model Creation and Evaluation:

   • Split the pre-processed dataset into training and testing sets.
   • Implement a RandomForestClassifier or GradientBoostingClassifier.
   • Perform hyperparameter tuning using GridSearchCV or RandomizedSearchCV.
   • Evaluate the model's performance using metrics such as accuracy and F1 score.

4. Report:

   • Answer specific questions related to the project in a separate document.
   • Summarize insights gained from EDA.
   • Explain the criteria for selecting columns to drop or keep.
   • Describe the hyperparameter tuning strategy used and its impact on model performance.
   • Present the final F1 score achieved by the model.

Progress Summary

EDA Highlights:

• Explored a dataset with over 600,000 transactions.
• No missing values identified.
• Diverse data types: float64, int64, and object.
• Discovered that 'isFlaggedFraud' column is indicative of poor fraud detection compared to actual fraud transactions.
• Found that most frauds occur in CASH_OUT and TRANSFER transactions.

Model Evaluation Results:

• Achieved 100% accuracy on the training set and 99.96% accuracy on the test set with a RandomForestClassifier.
• Observed a discrepancy in recall between fraudulent and non-fraudulent transactions in the test set.