🎭 Multi-Class Sentiment Analysis System

A production-grade, multi-class sentiment analysis system that classifies text into Positive, Negative, and Neutral sentiments using ensemble machine learning techniques.

Built as an academic AI/ML course project demonstrating mastery of NLP fundamentals, classification algorithms, and data science best practices.

---

🎯 Project Highlights

Feature	Description
3-Class Classification	Positive, Negative, Neutral sentiment detection
6 ML Algorithms	Logistic Regression, SVM, Naive Bayes, Random Forest, XGBoost, Voting Ensemble
SMOTE Oversampling	Handles class imbalance via synthetic minority oversampling
TF-IDF Features	Unigrams + Bigrams with sublinear TF weighting
10 Visualizations	Publication-quality dark-theme charts at 300 DPI
Real-Time Predictions	Streamlit web app with confidence scores
Feature Importance	Interpretable results showing top contributing words
Cross-Validation	5-fold stratified CV for robust evaluation

---

📂 Project Structure

Sentiment_Analysis/
├── Dataset/                          # Raw datasets (CSV, XLSX)
│   └── reddit_artist_posts_sentiment.csv   # Primary dataset (31K+ samples)
├── src/                              # Source code modules
│   ├── __init__.py
│   ├── utils.py                      # Configuration & utilities
│   ├── data_loading.py               # Data loading & EDA
│   ├── preprocessing.py              # NLP text preprocessing
│   ├── feature_engineering.py        # TF-IDF feature extraction
│   ├── model_training.py             # Model training with SMOTE
│   ├── evaluation.py                 # Metrics & model comparison
│   └── visualization.py             # Professional visualizations
├── outputs/
│   ├── figures/                      # Generated charts (PNG, 300 DPI)
│   └── models/                       # Saved trained models (joblib)
├── main.py                           # Master pipeline script
├── app.py                            # Streamlit web app
├── requirements.txt                  # Python dependencies
└── README.md                         # This file

---

🚀 Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run the Full Pipeline

python main.py

This will:

• Load and preprocess the dataset
• Train 6 ML models with SMOTE oversampling
• Evaluate all models with comprehensive metrics
• Generate 10 professional visualizations
• Save trained models for the web app

3. Launch the Web App

streamlit run app.py

---

📊 Models Implemented

#	Model	Key Configuration
1	Logistic Regression	C=1.0, balanced class weights, multinomial
2	Linear SVM	C=1.0, balanced weights, calibrated for probabilities
3	Multinomial Naive Bayes	Laplace smoothing (α=1.0)
4	Random Forest	200 trees, max_depth=50, balanced weights
5	XGBoost	200 estimators, learning_rate=0.1, max_depth=6
6	Voting Ensemble	Soft voting of all 5 base models

---

🔧 NLP Pipeline

1. Lowercasing — Normalize case
2. URL Removal — Strip http/https links
3. @Mention Removal — Remove Twitter-style handles
4. Hashtag Cleaning — Keep word, remove # symbol
5. HTML Entity Removal — Clean escaped characters
6. Repeated Character Normalization — "sooooo" → "soo"
7. Punctuation Removal — Strip special characters
8. Number Removal — Remove numeric tokens
9. Tokenization — Split into words
10. Stop Word Removal — NLTK English + domain-specific
11. Lemmatization — WordNet lemmatizer

---

📈 Visualizations

The system generates 10 publication-quality charts:

1. Class Distribution — Sentiment balance analysis
2. Text Length Distribution — Word/character counts per class
3. Word Clouds — Most frequent words per sentiment
4. Confusion Matrices — Normalized heatmaps for all models
5. Model Comparison — Grouped bar chart of all metrics
6. Cross-Validation Boxplot — CV score distributions
7. Feature Importance — Top 15 words per sentiment class
8. ROC Curves — Multi-class OvR for all models
9. Learning Curves — Train/validation accuracy vs. dataset size
10. Per-Class Performance — F1-score breakdown by class

---

🛠 Technology Stack

• Python 3.10+
• scikit-learn — ML models & evaluation
• XGBoost — Gradient boosting classifier
• NLTK — NLP preprocessing
• imbalanced-learn — SMOTE oversampling
• Matplotlib / Seaborn — Static visualizations
• Streamlit — Interactive web app
• Plotly — Interactive charts in web app

---

📝 License

Academic project — MIT License