Sentiment Classifier for Customer Feedback

A machine learning project that classifies customer feedback as positive or negative using natural language processing techniques. Built using Logistic Regression and deployed as an interactive Streamlit web app.

---

🗂️ Project Folder Structure

sentiment-classifier/
│
├── data/                       
│    ├── amazon_cells_labelled.txt
│    ├── imdb_labelled.txt
│    ├── readme.txt
│    └── yelp_labelled.txt
│
├── docs/
│    └── word-cloud.png
│
├── notebooks/                  
│    └── 01_eda_and_modeling.ipynb
│
├── models/
│    ├── logistic_model.pkl
│    └── tfidf_vectorizer.pkl
│                     
├── app/                      
│    └── app.py
│
├── CONTRIBUTING.md
├── LICENSE
├── README.md
└── requirements.txt

---

Overview

This project uses the UCI Sentiment Labelled Sentences dataset to classify short customer reviews from Yelp, Amazon, and IMDB. The final application lets users input a sentence and instantly receive a predicted sentiment with model confidence.

---

💡 Key Features

• Cleaned and explored labeled customer reviews (positive/negative)
• Trained and evaluated multiple models:
  • Logistic Regression ✅ (final model)
  • Multinomial Naive Bayes
  • Random Forest
• Visualized performance using bar plots
• Created word clouds for each sentiment class
• Built a simple and intuitive Streamlit app for real-time predictions

---

Model Performance

Model	Accuracy	Precision	Recall	F1 Score
Logistic Regression	0.836	0.837	0.818	0.828
Multinomial Naive Bayes	0.829	0.799	0.860	0.828
Random Forest	0.809	0.833	0.754	0.791

Logistic Regression was chosen for deployment due to its balanced and reliable performance across all metrics.

---

📦 Installation

1. Clone the repository:

git clone https://github.com/mandele1999/sentiment-classifier.git
cd sentiment-classifier

2. Install dependencies:

pip install -r requirements.txt

3. Run the app locally:

streamlit run app/app.py

Dataset

• Source: UCI Sentiment Labelled Sentences dataset
• Size: 2,748 sentences across Yelp, Amazon, and IMDB
• Labels: 1 (positive), 0 (negative)

Sample Predictions

Sentence: "Absolutely terrible experience, never buying again."
Predicted Sentiment: Negative (0.95 confidence)

Sentence: "Customer service was quick and helpful."
Predicted Sentiment: Positive (0.92 confidence)

Visual Insights

🌥️ Word Clouds

Positive Feedback	Negative Feedback
Frequently used words in positive reviews	Common terms in negative reviews

Project Summary

This project demonstrates how a simple yet powerful machine learning pipeline can be used to extract insights from customer feedback. It combines text preprocessing, model evaluation, and deployment—all in one place.

✨ Future Improvements

• Explore advanced models like DistilBERT or BERT
• Expand the dataset with more domain-specific reviews
• Add explanations for predictions using tools like LIME or SHAP
• Deploy on the web using Streamlit Community Cloud

Acknowledgements

Thanks to UCI for providing a clean and simple dataset to work with. The project was created as part of a data science portfolio to demonstrate NLP and deployment capabilities.