IMDb Sentiment Analysis with DistilBERT LoRA, TF-IDF Baselines, Calibration, Interpretability, Robustness Testing, and Semi-Supervised Learning

This tutorial demonstrates how to build a sentiment analysis system using movie reviews from the Stanford NLP IMDb dataset, comparing both classical machine learning approaches and modern transformer-based methods. The workflow includes establishing baseline performance with TF-IDF and Logistic Regression, then fine-tuning DistilBERT with LoRA for improved results, followed by evaluation through multiple metrics like accuracy, F1 score, and ROC-AUC. The tutorial emphasizes practical considerations often overlooked in sentiment analysis, including dataset auditing for class imbalance and duplicate reviews, probability calibration to ensure confidence scores are reliable, and robustness testing such as examining how truncating long reviews affects model performance. Finally, it demonstrates semi-supervised learning by using unlabeled reviews to create additional training data through confidence-based pseudo-labeling.