Fraud Detection Demonstration with MLOps

This demonstration presents a solution for financial fraud prevention using MLRun’s feature store to define complex features that help identify fraudulent activities. Fraud prevention is a challenging task as it requires processing transactions and events in real time, responding quickly to block transactions before they are completed.

Solution Approach

To address this problem, two pipelines are implemented:

• Development Pipeline: Allows testing and fine-tuning feature engineering logic and models.
• Production Pipeline: Uses the same features and models but is adapted to handle real-time data.

Additionally, data and model monitoring is automated, enabling the detection of deviations (drift) and triggering model retraining within a CI/CD pipeline. The complete process is illustrated in the following diagram:

Project Implementation Steps

Data Exploration and Analysis (EDA):

Understand the characteristics and structure of the data.

Construction of the Data Ingestion and Preparation Pipeline:

Preprocess and transform data for use in models.

Development of the Model Training and Validation Pipeline:

Train models using different features and algorithms.

Development of the Application Service Pipeline:

Intercept requests, process real-time data, and make inferences.

Data and Model Monitoring:

Detect deviations (drift) and evaluate model performance in production.

Continuous Operations and CI/CD Management:

Automate continuous integration and deployment to keep models up to date.