Skip to content

Model Development

challenges
Challenges in Dataset

Building the T/X Dataset

graph LR
    A[Data Generation] --> B[Generate Dataset]
    B --> C[Highly Imbalanced]
    B --> D[Contain both numerical & categorical values]
    B --> E[Features to be time dependent]
    B --> F[Data must contain fraud scenarios]

    A --> G[Baseline Prediction Model]
    G --> H[Feature Transformation]
    G --> I[Predictive Modelling]

    A --> J[Apply it on Real-world Dataset]
    J --> K[IEEE-CIS Dataset]

    A --> L[Transaction Features]
    L --> M[Customer ID]
    L --> N[Transaction ID]
    L --> O[T/X Amount]
    L --> P[Date & Time]
    L --> Q[Terminal ID]
    L --> R[Fraud Labels either 0 or 1]

    A --> S[Customer Profiles Generation]
    S --> T[Customer ID]
    S --> U[Geographical Location. Latitude & Longitude]
    S --> V[Mean Amount, Standard Deviation Amount]
    S --> W[Mean Normal T/X per day]

    A --> X[T/X Generation Process]
    X --> Y[Generation of terminal profiles]
    X --> Z[Association of customer profiles to terminals]
    X --> AA[Generation of customer profiles]
    AA --> AB[Difference in spending habits]
    AA --> AC[Their Geographical location]
    AA --> AD[Spending frequency]
    AA --> AE[Spending Amounts]
  • Generate a dataset having, 5,000 customers, 10,000 terminals and 183 days of transactions.
  • 1,754,155 transactions will generated in processing time about 3 minutes.

Addition of Fraud Scenarios

Any Transaction whose amount is more than 220 is a fraud.

Everyday 2 terminals drawn at random and keep tracking upto next 28 days and considering them fraudulent.

Everyday 3 customers are drawn at random.

  • In next 14 days, 1/3rd of their transactions have multiplied by 5 and marked as fraudulent.
  • All Card not present scenarios.
  • Note: Adding fraudulent scenarios about 14,681 Transactions. The ratio of fraudulent and non-fraudulent is 0.8% of all transactions.
  • This generated dataset shows around 10,000 transactions per day. The number of fraudulent transactions per day is around 85, which is a significant figure.

Feature Engineering

Date & Time features

Generating some more features and scenarios as shown below

  • Weekdays
  • Weekends
  • Day
  • Night

Customer ID features

Transforming Customer ID to characterize customer behavior

  • RFM (Recency, Frequency, Monetary) value.
  • Keep track of average spending amount & number of transactions.
  • Creating 6 new features based on their behavior on working days, non-working days, vacations, and tours, etc.

Method used for transactions

  1. Using cards:

    • Lost cards
    • Skimming the cards
    • Ordered a new card on the wrong address

  2. Without using cards:

    • Using Mobile/laptop and their versions
    • Using Net-banking / UPI / third-party apps
    • Loss of UPI PIN

Choosing the IEEE-CIS Dataset

  • IEEE-CIS Fraud Detection dataset is a large-scale dataset that contains transaction data from an e-commerce platform.
  • The dataset includes a wide range of features, such as transaction amount, product category, and customer information.
  • This meets the criteria for building a robust fraud detection model and merging the generated dataset with the REAL WORLD dataset.
  • The dataset contains 590,540 transaction details of 393 different features in the training dataset.

Working on the Dataset

  • After compiling the generated and IEEE-CIS dataset, there are many challenges in building the fraud detection model, such as:
    • Imbalance dataset between fraud and non-fraud dataset.
    • Dealing with a huge number of features.
    • Selection of correct and most important features.
    • Handling missing values.
    • Scaling the features.
model_flowchart
Model Flowchart