RetailPulse: RoBERTa-base review sentiment (v4.3)

Fine-tuned roberta-base (seed 42) that classifies a clothing review as negative, neutral or positive. It powers the RetailPulse Space, which routes reviews to customer service.

Data: Women's E-Commerce Clothing Reviews (Kaggle, public domain). Labels come from star ratings (1-2 negative, 3 neutral, 4-5 positive) and are noisy: in a 90-review single-annotator audit the author agreed with the star label on 66% of reviews (47% for neutral). Input is the review title and text.

Results (test set, model chosen and calibrated on validation only): macro-F1 0.706, negative recall 0.741. Routing policy: 95.5% of negative-star reviews reach a person, 1,253 staff reads per 5,000 reviews. Cross-fitted validation estimate: 1,289 reads/day, -20.2% illustrative cost vs Logistic Regression.

Calibration: temperature scaling, T = 1.4116; apply softmax(logits / T). The routing policy (thresholds, gates) is in retailpulse_policy.json.

Limitations: star-derived labels, English-only US retail reviews, single seed, and the test split was examined in earlier project versions. Illustrative business costs are assumptions, not observed retailer costs. Intended for a university assignment (BC3415), not for production decisions about customers.