Sentiment-analysis-with-hyper-parameter-tunning-and-model-evaluation-with-Xai

This project performs Sentiment Analysis on movie reviews using IMDb's large-scale dataset. It applies machine learning and NLP techniques to classify reviews as positive or negative, along with explainability (XAI) tools to interpret model decisions.

---

📌 Objective

• Convert IMDb ratings (1–10 scale) into binary sentiment classes:
  • 1–4 → Negative
  • 7–10 → Positive
• Build models to classify text-based reviews into positive or negative
• Address challenges like sarcasm, negation, and mixed sentiments
• Use Explainable AI (XAI) to understand and interpret the model’s predictions

---

📊 Dataset Overview

• 📁 Source: Kaggle IMDb Dataset
• 🔢 Size: 149,780 reviews
• 🎬 Columns:
  • Review: Text review
  • Rating: Numerical score (1–10)
  • Movie: Name of the movie
  • Resenhas: Portuguese review (dropped)

---

🧹 Data Preprocessing

• Expanded contractions (e.g., can't → cannot)
• Removed URLs, special characters, and irrelevant stopwords
• Retained negation words like “not” to preserve sentiment context
• Created a Review_clean column for processed text

---

📈 Exploratory Data Analysis (EDA)

• Verified class balance (equal reviews per rating)
• Plotted word clouds by sentiment
• Analyzed review length, word density, and average word length
• N-gram analysis (uni-, bi-, and trigrams) showed key sentiment patterns

---

🛠️ Feature Engineering

• Used CountVectorizer and TF-IDF for text-to-vector conversion
• Focused on unigrams, bigrams, and trigrams
• Applied chi-squared test for feature selection
• Found TF-IDF more effective in separating sentiment

---

🤖 Model Building & Evaluation

Models Used:

• Logistic Regression ✅ (Best performer)
• Decision Tree (Overfit)
• Random Forest (Some overfitting)

Metrics:

• Precision
• F1-score
• AUC-ROC

Best Model (Logistic Regression):

• Precision: 0.8955 (test)
• F1 Score: 0.8955
• AUC: 0.9606
• Fine-tuned using GridSearchCV

---

🚧 Challenges Faced

• Sarcasm: e.g., “Sure, best movie ever... I slept halfway.”
• Mixed polarity: Some reviews contained both praise and criticism.
• Negation Handling: “Not bad” ≠ “bad”

---

📌 Future Work

• Introduce deep learning models like LSTM or BERT
• Implement sarcasm detection modules
• Apply domain adaptation for different datasets (e.g., product reviews)

---

👥 Contributors

• Nikhil Gupta

---