The Culture: Ad Targeting CTR Model (XGBoost)
Binary classifier predicting whether an ad impression will be clicked. The dataset is severely imbalanced (~2.5% click rate), so the model uses scale_pos_weight and is evaluated primarily on PR-AUC, not accuracy or ROC-AUC alone.
Metrics
- PR-AUC: 0.0568
- ROC-AUC: 0.7548
- scale_pos_weight used: 39.00
Full classification report (default 0.5 threshold) and a threshold sweep are saved in eval_metrics.json / the training notebook. Because of the heavy imbalance, consider tuning the decision threshold for your use case (e.g. ranking ads by predicted probability rather than thresholding at 0.5).
Features
Categorical (one-hot): ['ad_category', 'ad_brand', 'target_archetype', 'target_age_range', 'target_gender', 'placement', 'user_archetype', 'user_age_range', 'user_gender', 'day_of_week', 'device_type'] Numeric: ['bid_amount_cents', 'user_session_number', 'hour'] Binary: ['archetype_match']
Note: was_converted was intentionally excluded as a feature since it's a downstream consequence of a click, not known at prediction time.
Usage
See the inference / Hugging Face call cells in the training notebook for a full working example.
Part of The Culture ML model suite.