Fraud Detection

Objective of the Project

The goal of the project is building a model using Machine learning that detects financial fraud by identifying suspicious activities.

Dataset Overview:

Fraud Detection Dataset
The dataset is divided into two parts: train and test.
It contains information about financial transactions, with a column indicating whether the transaction was fraudulent (target variable).

Key Variables:

• index: Unique Identifier for each row
• trans_date_trans_time: Transaction DateTime
• cc_num: Credit Card Number of Customer
• merchant: Merchant Name
• category: Category of Merchant
• amt: Amount of Transaction
• first: First Name of Credit Card Holder
• last: Last Name of Credit Card Holder
• gender: Gender of Credit Card Holder
• street: Street Address of Credit Card Holder
• city: City of Credit Card Holder
• state: State of Credit Card Holder
• zip: Zip of Credit Card Holder
• lat: Latitude Location of Credit Card Holder
• long: Longitude Location of Credit Card Holder
• city_pop: Credit Card Holder's City Population
• job: Job of Credit Card Holder
• dob: Date of Birth of Credit Card Holder
• trans_num: Transaction Number
• unix_time: UNIX Time of transaction
• merch_lat: Latitude Location of Merchant
• merch_long: Longitude Location of Merchant
• is_fraud: Fraud Flag <--- Target Class

Challenges & Solutions

• Multicollinearity between features - Solution: Addressed multicollinearity by combining features and eliminating unnecessary ones.
• Feature Selection - Solution: Kept features like ZIP, latitude, and longitude in the model, enhancing the feature engineering process for better fraud detection.

Methodology

ETL: Data Extraction & Cleaning

• Created a class named DataLoader that loads the training and test datasets from CSV files.
• Combined both datasets into one DataFrame (combined_train_test_data).
• Created a class named DataExplorer that performs initial exploration of the combined data, displaying: the shape and columns of the dataset, Data types of each column and basic information about the DataFrame and the first five rows for a quick overview.
• Created a class named DataLoader that renames the columns for lowercase, while dealing with missing values and duplicates and dropped unnecessary columns.

Exploratory Data Analysis, Data Preprocessing & Visualizations

• Created a class named EDA that provides an overview of the dataset, including descriptive statistics and data visualizations.
  • Statistical summary: Computes and returns a statistical summary for numerical columns, including measures of central tendency (mean, median) and dispersion (variance, standard deviation, interquartile range (IQR), and outlier counts).
  • Data Distribution Visualizations: Distribution of numerical features (excluding 'is_fraud') using histograms with KDE plots and distribution of the target variable 'is_fraud' using a count plot to compare fraudulent and non-fraudulent transactions.
  • Inferential Statistics: The Chi-square test was applied to each categorical feature against the target variable (is_fraud) to check the association with the target, categorical variables transformed for numeric and added to the correlation matrix.

Feature Engineering

• Created new features to the dataset to help address multicollinearity between existing features (time-related features, transaction amount features, and calculates the distance to the merchant).
• Additionally, features like time differences, small city indicators, and combined categories were generated, providing more diverse information, reducing the correlation between the original features and improving model performance.

Modelling

• Train-Test Split: Divided the data into 70% for training and 30% for testing.
• Scaling: Applied MinMaxScaler to normalize the numeric features in the training and test sets, scaling values between 0 and 1.
• SMOTE (Over-sampling Technique): Balanced the dataset by generating synthetic samples for the minority class (fraud) in the training data.

Models used:

• Logistic Regression: Performs well in predicting non-fraud cases with perfect precision but misses 31% of actual non-fraud cases (recall 0.69). However, it struggles with fraud detection, showing extremely low precision (0.01) and F1-score (0.02), making it unreliable for detecting fraud.
• Random Forest: Performs well on non-fraud cases with near-perfect precision and recall (0.99). While fraud detection improves over Logistic Regression, with 79% recall and a higher F1-score (0.51), its precision (0.38) remains low, limiting its reliability for fraud detection.
• XGBoost (model chosen): The model performs well in detecting non-fraudulent transactions, with 100% precision and recall. However, it only detects 81% of fraudulent transactions, missing 19% and having a 14% false positive rate. The high accuracy alone is not a reliable metric for this model due to the imbalanced data between the classes, where non-fraudulent transactions dominate.

Data Visualizations

• Feature Importance Plot: Used to check the impact on the model's decisions for refining and guiding feature engineering.
• Confusion Matrix plot: The confusion matrix was important for evaluating the model's performance, particularly in identifying false positives.
• ROC Curve plot: We concluded that the model is good at distinguishing between fraudulent and non-fraudulent transactions (performing with very few mistakes) and can correctly classify +90% of the cases based on its predictions.

Next Steps

Review the specific false positives and false negatives to understand the model’s mistakes and improve data preprocessing.

Deliverables

• Jupyter Notebook with Full Code in Python: Contains all steps from data preprocessing to model evaluation.
• PowerPoint Presentation: A concise overview of the project, including methodology, results, and key findings, supported by visual plots.
• Power BI Dashboard: An interactive visualization of key metrics and model performance, featuring insights from the table with real and predicted values to explore patterns, trends, and key insights.