THE PROBLEM THAT NEEDED SOLVING
Wanted to predict 'Return' events before they happen. Returns aren't just bad products—they're the gap between expectation (reviews) and reality (delivery).
WHAT WAS ACTUALLY IMPLEMENTED
Built Streamlit app comparing Logistic Regression, Random Forests, XGBoost on NLP features (TextBlob sentiment) + engineered metadata (Helpfulness Ratios).
Model Evaluation & Performance Analysis
Built a comparative framework to identify the optimal predictor for high-variance e-commerce data.
Feature Engineering: The 'Helpfulness' Gap
Standard review analysis misses a key signal: how others perceive the review. High negative sentiment paired with a high 'helpfulness' count is a 94% certain predictor of a return event.
High Risk Scenarios
Model Breakthrough: XGBoost Capture non-linear interactions between delivery delays and review polarity. Logistic Regression failed here, but XGBoost achieved a peak AUC of 0.89.
Low Risk Baseline
Validated the model against 'Happy Path' deliveries—identifying customers who consistently keep products despite neutral reviews.