Business Context

A Google Research Scientist candidate is asked to reason through a realistic first-round ML problem rather than recite isolated theory. Build a lightweight classifier for YouTube comment spam detection and compare a classical ML baseline with a simple neural baseline, explaining the statistical and modeling tradeoffs clearly.

Dataset

Use a historical moderation dataset from YouTube comments collected across popular channels. Each row is one comment with metadata available at prediction time.

• Size: 120K comments, 12 engineered non-text features plus raw text
• Target: Binary label — spam (1) vs non-spam (0)
• Class balance: 18% spam, 82% non-spam
• Missing data: ~8% missing in account_age_days and language; some comments are empty after text cleaning

Success Criteria

A good solution should outperform a majority-class baseline and deliver strong ranking quality for moderation triage. Aim for AUC-ROC > 0.92, F1 > 0.78, and recall > 0.85 at precision >= 0.75 on a held-out test set.

Constraints

• Inference should support near-real-time moderation in <20 ms per comment in a Google production setting
• The solution should be interpretable enough to explain obvious spam signals to Trust & Safety reviewers
• Retraining can run daily; serving cost should remain modest

Deliverables

1. Train a classical baseline model and one simple deep learning baseline
2. Justify feature preprocessing, regularization, and validation strategy
3. Evaluate with threshold-free and threshold-based metrics
4. Explain how you would choose the operating threshold for moderation
5. Describe failure modes, including language drift and adversarial spam patterns