Adaptive Graph-Enhanced Fraud Detection Platform
A streaming intelligence system for detecting suspicious financial activity using transaction-level and graph-based behavioural signals.
Project Overview
The Adaptive Graph-Enhanced Fraud Detection Platform is designed to identify suspicious financial activity while transactions are still moving through the payment stream.
Traditional fraud-detection models often evaluate transactions independently. This platform extends that approach by analysing relationships between accounts, devices, merchants, and transaction histories. These connections provide additional behavioural context that can reveal coordinated fraud patterns.
Apache Kafka handles real-time transaction ingestion, while Spark Structured Streaming performs continuous validation, transformation, and feature generation. The resulting transaction-level and graph-based features are passed to a machine-learning detection layer that estimates the risk associated with each transaction.
The project demonstrates the integration of streaming data engineering, graph-based analysis, machine learning, and scalable fraud-monitoring architecture.
Technology Stack
- Python
- Apache Kafka
- Apache Spark Structured Streaming
- Machine Learning
- Graph-Based Feature Engineering
- Streaming Data Processing
System Architecture
The platform uses a layered streaming architecture that transforms incoming financial transactions into enriched behavioural features before producing a fraud-risk prediction.
- Transaction Producer — receives or generates incoming financial transaction events.
- Apache Kafka — handles real-time transaction ingestion and message delivery.
- Spark Structured Streaming — continuously validates, transforms, and processes transaction data.
- Graph Feature Engine — models relationships between accounts, devices, merchants, and transactions.
- Fraud Detection Model — estimates the fraud risk associated with each transaction.
- Alert and Monitoring Layer — surfaces suspicious activity for investigation.
System Workflow
- A financial transaction enters the platform as a streaming event.
- Apache Kafka receives the event and publishes it to the transaction stream.
- Spark Structured Streaming cleans, validates, and transforms the transaction data.
- Transaction-level features such as amount, timing, frequency, and location are generated.
- Graph relationships are analysed to identify connected accounts, shared devices, repeated merchants, and coordinated activity.
- The machine-learning model combines transactional and graph-derived features.
- The transaction is classified as legitimate or potentially fraudulent.
- High-risk transactions are forwarded to the alert and monitoring layer.
Key Features
- Continuous processing of streaming financial transactions.
- Graph-based analysis of relationships between financial entities.
- Combination of behavioural, transactional, and relational signals.
- Detection of coordinated and evolving fraud patterns.
- Scalable processing using Kafka and Spark.
- Real-time fraud-risk scoring and alert generation.
Results
- Built an end-to-end streaming fraud-detection workflow.
- Combined transaction-level features with graph-based behavioural information.
- Supported low-latency processing of continuously arriving transactions.
- Detected suspicious activity that may not be visible from isolated transactions.
- Created a scalable foundation for real-time financial fraud monitoring.