Business Context

ShopLens, an e-commerce analytics company, wants a baseline text classification pipeline to automatically label customer reviews as positive or negative before routing low-rated feedback to support. The hiring team wants to assess your practical knowledge of Pandas for data preparation, Scikit-Learn for classical ML, and PyTorch or TensorFlow for a simple neural baseline.

Dataset

You are given a historical review dataset exported from the marketplace data warehouse.

• Size: 48K reviews, 8 input features
• Target: Binary sentiment label: positive (1) vs negative (0)
• Class balance: Moderately imbalanced — 68% positive, 32% negative
• Missing data: 12% missing in helpful_votes, 7% missing in device_type, 3% empty review_text

Success Criteria

A strong solution should achieve F1 >= 0.84 on the negative-review class and ROC-AUC >= 0.90 on the held-out test set. The candidate should also compare at least one classical model with one neural-network approach and explain tradeoffs.

Constraints

• Batch scoring only; no real-time serving requirement
• Training should complete on a standard laptop or single CPU/GPU instance
• The solution must be maintainable by a small data team and easy to retrain weekly
• Some interpretability is required for business stakeholders

Deliverables

1. Build a complete preprocessing pipeline using Pandas and appropriate imputations.
2. Train a Scikit-Learn baseline model for sentiment classification.
3. Train either a PyTorch or TensorFlow neural model on the same task.
4. Compare model quality, training complexity, and deployment tradeoffs.
5. Report final metrics on a held-out test set and explain which approach you would ship first.