Business Context

Google Ads wants to predict whether a newly created search ad will receive a low quality rating within 7 days, so policy and optimization systems can intervene early. You are given a supervised learning dataset and asked to design a feature selection strategy that improves model quality without introducing leakage or making the model too expensive to serve.

Dataset

The training data contains ad-level snapshots collected at ad creation time and aggregated over the first 24 hours only.

• Rows: 420K ads from the last 9 months
• Target: low_quality_7d = 1 if the ad is rated low quality within 7 days, else 0
• Class balance: 11.5% positive, 88.5% negative
• Missing data: ~9% missing in landing page features, ~4% missing in early performance for low-traffic ads

Success Criteria

A good solution should improve validation PR-AUC over a simple all-features logistic regression baseline by at least 10%, while keeping offline feature generation simple enough for daily retraining and online scoring under 20 ms per ad.

Constraints

• Avoid target leakage from post-label information
• Prefer features that can be computed consistently in Google Ads production pipelines
• The final model must remain interpretable enough for QA and policy review
• Feature computation cost matters more than squeezing out tiny offline gains

Deliverables

1. Propose a feature selection framework for this supervised classification problem.
2. Build a baseline model and a reduced-feature model.
3. Justify which features to keep, transform, or remove.
4. Show how you would validate feature usefulness without leakage.
5. Report evaluation metrics and explain tradeoffs between performance, interpretability, and serving cost.