📝 Text Classification using Machine Learning & Deep Learning

This repository demonstrates a complete Natural Language Processing (NLP) pipeline for text classification, progressing from a traditional machine learning baseline to a deep learning model built with TensorFlow and Keras.

The project is designed to be educational, extensible, and presentation-ready, making it suitable for learning, interviews, and portfolio showcasing.

---

📌 Problem Statement

Given a piece of text, the task is to classify it into one of several predefined categories using supervised learning techniques.

---

📂 Dataset

• Dataset: 20 Newsgroups (via scikit-learn)
• Selected Categories:
  • rec.autos
  • sci.med
  • comp.graphics
  • sci.space
• Total Samples: ~3,940 documents
• Preprocessing:
  • Removed headers, footers, and quoted text
  • Used cleaned raw text for modeling

---

🧠 Project Overview

This project is implemented in two stages to clearly compare traditional NLP methods with deep learning approaches.

---

1️⃣ TF-IDF + Logistic Regression (Baseline)

📓 Notebook: 01_tfidf_logistic_regression.ipynb

Methodology

• Text Representation: TF-IDF Vectorization
  • Max features: 5000
  • English stopwords removed
• Model: Logistic Regression
• Train–Test Split: 80% / 20%

Evaluation

• Accuracy
• Precision, Recall, F1-score

Results

• Test Accuracy: ~88%
• Strong, interpretable baseline with minimal computation

---

2️⃣ Deep Learning with TensorFlow & Keras

📓 Notebook: 02_tensorflow_text_classification.ipynb

This notebook replaces manual feature engineering with learned word embeddings and a neural network architecture.

Model Architecture

• TextVectorization layer
• Embedding layer
• Global Average Pooling
• Dense hidden layer (ReLU)
• Softmax output layer

Training Details

• Framework: TensorFlow & Keras
• Loss Function: Sparse Categorical Crossentropy
• Optimizer: Adam
• Epochs: 10
• Batch Size: 32
• Validation Split: 20%

Results

• Validation Accuracy: ~81–85%
• Test Accuracy: ~81%
• Demonstrates semantic learning and scalability

---

📊 Model Comparison

Model	Feature Type	Test Accuracy
TF-IDF + Logistic Regression	Sparse statistical features	~88%
TensorFlow Neural Network	Learned embeddings	~81%

---

✅ Key Takeaways

• TF-IDF with linear models remains a strong NLP baseline
• Deep learning models learn semantic representations automatically
• Neural networks scale better for large and complex datasets
• The project highlights trade-offs between interpretability and flexibility

---

🚀 Future Improvements

• CNN-based text classifiers
• LSTM / Bi-LSTM architectures
• Transformer models (BERT)
• Hyperparameter tuning
• Confusion matrix & detailed error analysis

---

🧪 Requirements

• Python 3.x
• scikit-learn
• TensorFlow
• NumPy
• Jupyter Notebook / Google Colab

---

📜 License

This project is licensed under the MIT License.