Design a Fraud Detection System for Reviews/Ratings

To design a fraud detection system for Amazon reviews, I would start by defining our goal: minimizing false positives while catching coordinated manipulation in near real-time. First, we ingest clickstream and review data into Kafka, splitting it into a fast path for immediate alerts and a slow path for deep analysis. For the behavioral layer, we calculate features like time-to-first-review after account creation, review length variance, and sentiment deviation from product norms. Crucially, we implement a network graph service where nodes represent users, products, and IPs, and edges denote interactions. We run a sliding window algorithm to detect dense subgraphs indicating bot farms, looking for clusters where multiple accounts review the same item within seconds. Using Graph Neural Networks, we assign risk scores based on topological anomalies. If a score exceeds a dynamic threshold, the review is flagged for quarantine rather than immediate deletion to preserve trust. We also integrate a feedback loop where human analysts validate samples, retraining the model weekly. This approach ensures we protect the marketplace without disrupting genuine customer voices, aligning with Amazon's focus on long-term customer trust over short-term engagement metrics.