Analyzed 194K Amazon reviews to test whether star ratings reflect actual text sentiment, and which model best detects negative/neutral feedback despite 77% positive imbalance. Built 4 models: baseline & class-weighted Logistic Regression, Naive Bayes, and XGBoost. Evaluated on macro F1 and minority-class recall, not just accuracy. Key finding: Balanced Logistic Regression won baseline model hit 85% accuracy but missed 73% of neutral reviews. XGBoost's added complexity didn't beat the simpler linear model on sparse text features.