• Version
• Download 4
• File Size 355.77 KB
• Download

Context-Aware Fraud Detection Models Deployed via Feature Pipelines in ML Flow

Ravi Kiran Alluri

ravikiran.alluirs@gmail.com

Abstract- The ability and proficiency of fraud perpetrators to deceive and manipulate financial and digital systems have necessitated the development of intelligent and context-aware fraud detection models. Classical fraud detection systems are characterized by static rule-based systems and basic machine learning approaches that do not consider real-time context signals as user activity, transaction speed, device location, and session veracity. To overcome these shortcomings, this paper presents a process for creating and deploying context-aware fraud detection models non-intrusively, based on feature pipelines with MLFlow integration.

The proposed solution focuses on dynamic feature engineering, utilizing features such as behavior patterns, space-time information, and device-level fingerprints. The context-rich features are consumed by modular pipelines, which make it easy to transform, validate, and reuse in various machine learning workflows. Using MLFlow for experiment tracking, model versioning, and automatic deployment, the system provides traceability, reproducibility, and manageability along the life cycle of fraud detection models.

This paper demonstrates how contextual features can increase the average ROI when used in conjunction with XGBoost, LightGBM, Random Forest, and other popular machine learning algorithms. A comparative study is performed on a structured transaction dataset enriched with surrounding information. Standard evaluation metrics (precision, recall, F1-score) show significant enhancements over the non-contextual baseline models. Furthermore, the framework includes monitoring and automatic drift detection with its integrated features of MLFlow so that models can accommodate the latest fraud techniques (about 1 hour after) in near real-time.

The primary objective of this work is to leverage context-aware fraud detection models in conjunction with MLFlow's operational features. This enables data scientists and fraud analysts to collaborate on creating, testing, and deploying models to production, ensuring they are both explainable and compliant with regulatory standards. The paper also describes the architecture's ability to provide continuous model improvement through reinforcement loops with live environments.

The context-aware method proposed in this paper significantly enhances the accuracy of fraud detection and reduces false positives. Paired with an MLFlow-heavy MLOps infrastructure, this approach provides a scalable and transparent solution for both financial institutions and digital platforms seeking to strengthen their defenses against emerging forms of fraudulent behavior. This study opens up new prospects for fraud analytics research by connecting the dots between contextual intelligence and “productionalised” deployment processes.

Keywords- Context-aware fraud detection, MLFlow, feature pipelines, MLOps, machine learning, real-time analytics, financial fraud, model deployment, behavioral analysis, data lineage, feature engineering, model drift, adaptive fraud detection, transactional intelligence.