💬 Social Media Sentiment Analysis

An end‑to‑end NLP project for analyzing sentiment in social media posts (positive / negative / neutral).
Includes data cleaning, text normalization, feature extraction (TF‑IDF / embeddings), classical ML baselines (Logistic Regression, Naive Bayes, SVM), and optional deep learning (Bi‑LSTM). Provides rigorous evaluation with Accuracy, F1‑score, and confusion matrix.

---

📖 Overview

This repository demonstrates a reproducible sentiment analysis pipeline, from raw text to deployable models.
It covers:

• Preprocessing & normalization (lowercasing, punctuation removal, stopwords, lemmatization)
• Feature extraction with TF‑IDF (n‑grams) or pretrained embeddings
• Model training & comparison (LogReg, NB, SVM, Random Forest)
• (Optional) Bi‑LSTM sequence model for improved context handling
• Explainability & error analysis (misclassification review)

---

🗂️ Dataset

• Input: CSV with columns like: id, text, label (pos/neg/neu or 1/0).
• Location: place your files under Dataset/ (e.g., train.csv, test.csv).
• Class balance: check for imbalance; consider stratified splits and class weights.

If you use a public dataset (e.g., Twitter Sentiment, Sentiment140), cite the source in this README.

---

🧹 Preprocessing

• Clean text: lowercase, remove URLs, mentions, hashtags (optional keep hashtag text), emojis handling
• Tokenization, stopword removal, lemmatization
• TF‑IDF with n‑grams (1–2 or 1–3), max_features cap
• (Optional) Emoji/emoticon normalization and slang expansion

---

🧠 Models

• Logistic Regression (strong linear baseline on TF‑IDF)
• Multinomial Naive Bayes (fast baseline for sparse text)
• Linear SVM (robust with TF‑IDF)
• Random Forest (tabular baseline)
• (Optional) Bi‑LSTM with pretrained word embeddings

---

📈 Evaluation (replace with your numbers)

Model	Accuracy	F1 (macro)
Logistic Regression	0.90	0.89
Multinomial NB	0.88	0.87
Linear SVM	0.92	0.91
Random Forest	0.86	0.85
Bi‑LSTM (optional)	0.93	0.92

Export:

• Confusion matrix for 3‑class sentiment
• Per‑class precision/recall/F1
• Top informative features (for linear models)

---

🧩 Repository Structure (suggested)

Social-Media-Sentiment-Analysis/
├─ Dataset/                      # train/test CSVs
├─ Notebook/
│  └─ Sentiment_Analysis.ipynb   # main notebook
├─ src/                          # optional scripts
│  ├─ data.py                    # loading/cleaning
│  ├─ features.py                # TF-IDF / embeddings
│  ├─ train.py                   # training & tuning
│  └─ eval.py                    # metrics & plots
├─ reports/figures/              # CM, PR curves, feature plots
├─ requirements.txt
├─ .gitignore
└─ README.md

---

⚙️ Setup & Usage

1. Clone & install

git clone https://github.com/ziaee-mohammad/Social-Media-Sentiment-Analysis.git
cd Social-Media-Sentiment-Analysis
pip install -r requirements.txt

2. Run notebook

jupyter notebook Notebook/Sentiment_Analysis.ipynb

3. (Optional) Run scripts

python -m src.train --model "svm" --ngrams 1,2 --max_features 200000
python -m src.eval  --report

---

📦 Requirements (example)

pandas
numpy
scikit-learn
nltk
matplotlib
seaborn

If using Bi‑LSTM: add torch (or tensorflow) and suitable tokenizers/embeddings.

---

✅ Good Practices

• Use stratified train/test split
• Keep vectorizer + model in a single Pipeline to avoid leakage
• Fix random seeds for reproducibility
• Save vectorizer/model artifacts if you plan to deploy

---

🏷 Tags

data-science
machine-learning
nlp
sentiment-analysis
text-mining
classification
python
scikit-learn
tf-idf
``

---

## 👤 Author
**Mohammad Ziaee** — Computer Engineer | AI & Data Science  
📧 [email protected]  
🔗 https://github.com/ziaee-mohammad

---

## 📜 License
MIT — free to use and adapt with attribution.