Bank Fraud Detection

Data Science Project

Introduction

Bank fraud is a critical challenge in the financial sector, demanding innovative solutions for timely detection. Leveraging Neural Networks for this purpose offers a promising avenue. Neural Networks excel at recognizing intricate patterns, making them well-suited to identify fraudulent activities in the dynamic landscape of financial transactions.

Dataset Description

• We are using a dataset from Kaggle but it got 1,000,000 row. So, we take only 20,000 row from it to be lighter for the resources to run it.
• Here is the columns description:
• fraud_bool (Target Column): Binary indicator (0 or 1) representing whether the transaction is fraudulent or not.
• income: The reported income of the customer.
• name_email_similarity: A measure of similarity between the customer's name and email.

​

Dataset Description (cont.)

• prev_address_months_count: Number of months since the customer's last address change.
• current_address_months_count: Number of months at the current address.
• customer_age:Age of the customer.
• days_since_request: Number of days since the customer's request.
• intended_balcon_amount: The intended loan or credit amount.
• payment_type: The type of payment used for the transaction.
• zip_count_4w:Count of transactions within the last 4 weeks in the customer's ZIP code

Dataset Description (cont.)

• velocity_6h: Transaction velocity in the last 6 hours.
• velocity_24h: Transaction velocity in the last 24 hours.
• velocity_4w: Transaction velocity in the last 4 weeks.
• bank_branch_count_8w: Count of bank branches involved in transactions in the last 8 weeks.
• date_of_birth_distinct_emails_4w: Count of distinct email addresses associated with the customer's date of birth in the last 4 weeks.
• employment_status: The employment status of the customer.
• credit_risk_score: Credit risk score assigned to the customer.
• email_is_free: Binary indicator (0 or 1) denoting whether the email used is a free service.

​

Dataset Description (cont.)

• housing_status: The housing status of the customer.
• phone_home_valid: Binary indicator (0 or 1) indicating the validity of the home phone number.
• phone_mobile_valid: Binary indicator (0 or 1) indicating the validity of the mobile phone number.
• bank_months_count: Number of months the customer has been associated with the bank.
• has_other_cards: Binary indicator (0 or 1) representing whether the customer has other credit cards.
• proposed_credit_limit: The proposed credit limit for the customer.
• foreign_request: Binary indicator (0 or 1) indicating whether the transaction request is from a foreign source.

​

Dataset Description (cont.)

• source: The source of the transaction.
• session_length_in_minutes: Duration of the transaction session in minutes.
• device_os: The operating system of the device used for the transaction.
• keep_alive_session: Binary indicator (0 or 1) denoting whether the session is kept alive.
• device_distinct_emails_8w: Count of distinct email addresses associated with the device in the last 8 weeks.
• device_fraud_count: Count of fraud-related incidents associated with the device.
• month: The month in which the transaction occurred.

Preprocessing

Preprocessing (cont.)

• We have removed column “prev_address_months_count” because most rows are missing values “-1”
• We have replaced values of “-1” in column : “session_length_in_minutes” with the mean of the whole column

Preprocessing (cont.)

• We have replaced “-1” values in column “device_distinct_emails_8w” with “1”
• Replaced “-1” values in column “current_address_months_count” with mean of the column’s value

Preprocessing (cont.)

• Replaced “-1” values in column “bank_months_count” with mean of the column’s value

​

Preprocessing (cont.)

• We have done Label Encoding where we converted all categorical data to numeric values

Preprocessing (cont.)

• We have used MinMax to to compute the minimum and maximum values of the features and then scale the features within the range [0, 1] to normalize it to have better scaling.

Visualization

Histogram for representing the amount of fraud vs non fraud transactions �(and they are balanced)

Models

We implemented 2 models from scratch �& used 1 pre-trained model (TabNet).

First Model

First Model

Both training accuracy and validation accuracy are performing well during the model fitting

​

NO Underfitting

First Model Results

Accuracy : 94.01%

Other metrics

First Model Results

Moderate Overfitting

Model Modification

Model Modification

Model Modification Results

Model Modification Results

Moderate Overfitting

TabNet Model

TabNet Model

TabNet Model