Threshold Confidence
Ferretly determines threshold confidence for flagging posts through a dynamic, sophisticated framework designed to balance precision and recall in detecting specific risky behaviors. Our machine learning models are trained on millions of human-reviewed posts, capturing nuanced patterns in text, images, and context, and we periodically retrain them incorporating ongoing analyst feedback to adapt to evolving language and trends. Rather than applying a fixed confidence score, the effective threshold is dynamically adjusted based on a risk-minimization strategy that prioritizes minimizing both false positives (to avoid unnecessary reviews) and false negatives (to ensure critical risks are not overlooked). Complementing the ML predictions, Ferretly leverages whitelisted keywords—including extensive slang, street terms, and euphemisms—for behaviors like drug mentions; posts matching these keywords often register below the primary ML confidence threshold but are escalated for further upstream review. This secondary layer combines advanced large language models for deeper contextual analysis with human adjudication, ensuring comprehensive and accurate flagging while maintaining high reliability in our social media screening process.