Business Context

ShopNow, a mid-sized e-commerce marketplace, wants to predict whether a product listing will convert within 7 days of first impression so the ranking team can prioritize high-potential items. You must decide whether a classical ML model or a deep learning model is the better production choice for this dataset and business setting.

Dataset

The training data contains listing-level examples collected over 12 months.

• Size: 420K listings, 38 raw features
• Target: Binary — converted within 7 days (1) vs not converted (0)
• Class balance: 14% positive, 86% negative
• Missing data: ~9% missing in brand and seller metadata, ~18% missing in short descriptions for long-tail sellers

Success Criteria

A good solution should improve over a logistic regression baseline by at least +0.05 ROC-AUC and justify the model choice using data size, feature modality, latency, and maintainability. The final recommendation should be defensible for a production ranking pipeline.

Constraints

• Online inference budget: p95 < 20 ms per item
• Model must score millions of listings daily
• Ranking team needs some interpretability for debugging
• Retraining can run daily, but GPU budget is limited

Deliverables

1. Build a strong classical ML baseline and a deep learning candidate.
2. Compare them using offline metrics, latency, and operational complexity.
3. Explain when deep learning is justified for this dataset and when it is not.
4. Propose a production recommendation, including fallback criteria.
5. Describe how you would monitor model quality after deployment.