Sentiment analysis is crucial for understanding learner perceptions in online education. Existing approaches often rely on single sentiment labels, limiting their ability to capture nuanced feedback. This study introduces a hybrid framework that integrates rating-based, text-based, and combined sentiment labels, along with feature extraction using TF-IDF, GloVe, and FastText. Multiple machine learning and ensemble models were optimized through GridSearchCV and RandomizedSearchCV. Findings reveal that TF-IDF features consistently outperformed embeddings, while ensembles such as Random Forest, Gradient Boosting, and XGBoost delivered the best results. The hybrid label configuration achieved 84.62% accuracy with balanced precision, recall, and F1 scores. Interpretability was enhanced using confusion matrices, ROC curves, and LIME visualizations, which clarified feature contributions to sentiment predictions. Overall, the proposed framework improves accuracy, transparency, and scalability, with potential applications in e-learning, e-commerce, and social media monitoring