The system separates social context, experience engagement, account activity, communication patterns and learned user representations. XGBoost classifiers, a multilingual transformer and compact neural networks each produce age-band probabilities, which a calibrated gradient-boosted meta-learner combines.
Training uses nearly 30 million high-confidence labels, but Roblox says verified users are not representative of everyone. It reweights samples with survey-informed population targets, filters questionable labels, stores derived features instead of raw activity where possible, and uses the same data contracts for training and inference.