🤖 Sentiment Analysis Using TF-IDF and a Neural Network

📝 Project Overview

This project focuses on binary sentiment classification using book reviews. Given the text of a review, the goal is to predict whether the sentiment is positive or negative. This has been implemented by training a feedforward neural network for improved performance.

This work is part of my learning journey through the Break Through Tech AI Program, where I applied machine learning concepts in natural language processing (NLP).

---

📚 What the Project Covers

✅ 1. Define the ML Problem

• Problem Type: Binary classification
• Input: Raw text from book reviews
• Output: Sentiment label (positive or negative)

🧹 2. Data Preprocessing

• Tokenized and cleaned text reviews

• Removed stopwords and punctuation

• Transformed text using TfidfVectorizer with:

  • Custom max_df and min_df thresholds
  • Max features set to 3000

🤖 3. Model Development

• Model Type: Feedforward Neural Network using tensorflow.keras

• Architecture:

  • Dense hidden layers with ReLU activation
  • Dropout regularization
  • Final sigmoid layer for binary output

• Trained on the TF-IDF transformed text data

• Used binary crossentropy loss and SGD optimizer

💾 4. Model Persistence

• Saved both the trained model and the vectorizer using:

  • model.save() from TensorFlow
  • pickle.dump() for the TF-IDF vectorizer

📈 5. Evaluation

• Accuracy, precision, recall, and F1-score calculated on test data
• Plotted the precision-recall curve to analyze classifier performance

---

🔧 Technologies Used

• Python
• pandas, numpy
• scikit-learn (TfidfVectorizer, metrics)
• TensorFlow / Keras (modeling)
• matplotlib, seaborn (visualizations)
• pickle (persistence)

---

🎯 Key Takeaways

• Implemented a custom neural network for NLP tasks
• Tuned TF-IDF vectorizer parameters to improve model generalization
• Understood the importance of precision-recall trade-offs
• Learned to persist and reload models and vectorizers for deployment use cases

---

📌 How to Reuse the Model

• Load the saved TF-IDF vectorizer from the .pkl file
• Load the neural network using tensorflow.keras.models.load_model()
• Pass new text data through the vectorizer, then predict with the model

---

🤔 Reflection

This project helped me solidify my understanding of:

• NLP text vectorization techniques
• Neural network design and optimization
• Model persistence for real-world ML pipelines
• Evaluation strategies beyond accuracy