Fraud Prediction Hackathon Report

📝 Overview

This project addresses the challenge of identifying fraudulent transactions using a rich dataset composed of card transactions, user information, card metadata, and merchant category codes (MCC). The goal is to train a robust machine learning model capable of predicting fraud on unseen transactions.

📁 Datasets Used

• transactions_train.csv: Transaction-level features for training
• train_fraud_labels.json: Target labels (Yes or No) for fraud
• evaluation_features.csv: Transactions to be predicted (test set)
• cards_data.csv: Card metadata (brand, chip status, limits)
• users_data.csv: User demographics (income, location, credit)
• mcc_codes.json: Merchant category codes with descriptions

🎯 Objective

To accurately classify transactions as fraudulent (1) or not (0) and submit predictions in the correct format (transaction_id,fraud).

⚙️ Pipeline Summary

1. Data Loading & Merging

• Loaded all datasets using pandas.
• Converted mcc_codes.json to a structured DataFrame.
• Merged transactions with fraud labels, card info, user info, and MCC descriptions.
• Added a train flag to distinguish training and evaluation data.

2. Preprocessing

• Converted fraud labels Yes/No to binary (1/0).
• Handled missing values and assessed column completeness.
• Cleaned categorical features and explored value distributions (e.g. use_chip, merchant_state).
• Combined datasets using appropriate keys (client_id, card_id, mcc).

3. Exploratory Data Analysis (EDA)

• Assessed class imbalance: Only ~0.15% of transactions were fraudulent.
• Identified high-cardinality and dominant classes in object fields.
• Evaluated missing value patterns and planned imputation strategies.

4. Feature Engineering

• Parsed dates to extract time-based features.
• Encoded categorical variables.
• Cleaned numerical columns (e.g., removed $ in amount, converted to float).

5. Modeling (to be completed/added if done)

• Train-test split with stratification.
• Model candidates: Logistic Regression, Random Forest, XGBoost, etc.
• Evaluation metrics: Precision, Recall, F1-score due to class imbalance.

6. Prediction & Submission

• Generated predictions for the test set.
• Formatted submission file: submission.csv with columns:

Author:TN@ TP
Date: 15/05/2025

Built With

• python