NLP Sentiment Analysis — Amazon Reviews

Multinomial Naive Bayes + TF-IDF

Overview

This project builds a baseline sentiment classification system to automatically label Amazon product reviews as negative (0) or positive (1) .

The primary objective is business-driven : to help teams (Customer Service, Product, Operations) identify negative reviews faster , reduce manual triage time, and prioritize customer issues more effectively.

The solution uses TF-IDF features and a Multinomial Naive Bayes model, chosen for its speed, interpretability, and ease of deployment.

---

Key Highlights

• Dataset : Amazon Reviews (Kaggle)
• Task : Binary sentiment classification
• Negative: ratings 1–2
• Positive: ratings 4–5
• Neutral (rating 3) is excluded
• Text Representation : TF-IDF (unigram + bigram)
• Model : Multinomial Naive Bayes
• Evaluation Focus : Performance on negative reviews
• Threshold-based decisioning for operational flexibility

---

Model Performance (Test Set)

Using the best configuration (alpha = 2.0) and default threshold thr_pos = 0.50:

• Accuracy : ~0.86
• Negative class
  • Precision_neg ≈ 0.88
  • Recall_neg ≈ 0.85
  • F1_neg ≈ 0.86
• ROC AUC : ≈ 0.94

Threshold tuning allows the model to trade off between:

• catching more customer complaints ( higher recall ), or
• keeping review flags cleaner ( higher precision ).

---

Business Impact (Illustrative Simulation)

Assuming 20 seconds to manually review one customer comment:

• At thr_pos = 0.55 :
  • ~1,100 reviews automatically flagged as negative
  • Recall_neg ≈ 0.91
  • Estimated ~6 hours of manual work saved per 2,000 reviews

This demonstrates that even a simple model can deliver immediate operational value when paired with threshold-based decisioning.

---

Data Usage Clarification (EDA vs Modeling)

The same source dataset (Reviews.csv) is used in two distinct ways:

• Exploratory Data Analysis (EDA) Conducted on the full dataset to understand:
  • rating distribution,
  • review volume trends,
  • real-world imbalance toward positive reviews.
• Modeling & Evaluation Performed on a balanced subset sampled from the same dataset to:
  • ensure fair evaluation between negative and positive classes,
  • speed up experimentation,
  • build a clean and interpretable baseline.

This separation is intentional and documented in the notebook.

---

Project Structure

• nlp_nb_amazon_reviews.ipynb — end-to-end analysis and modeling notebook
• dataset/ — raw and sampled review data
• artifacts_nb/ — saved TF-IDF vectorizer and trained model
• README.md — project summary and usage notes

---

How to Run

1. Clone the repository

   git clone https://github.com/your-username/nlp-sentiment-analysis.git
   cd nlp-sentiment-analysis

2. Install dependencies (Python ≥ 3.9 recommended) :
   • pandas
   • numpy
   • scikit-learn
   • matplotlib
   • nltk
   • wordcloud
3. Open and run the notebook:

   jupyter notebook nlp_nb_amazon_reviews.ipynb

---

Limitations

• The model struggles with negation and sarcasm (e.g., “not bad at all” , “great… if you like disappointment” ).
• Sensitive to language drift and new product terminology.
• Interpretability is word-based , not instance-level (e.g., SHAP/LIME).

---

Future Improvements

• Enhanced preprocessing (negation handling, trigram features)
• Model comparison with Logistic Regression and Linear SVM
• Monitoring dashboard for Precision/Recall of negative reviews
• Evaluation of Transformer-based models (DistilBERT / BERT)
• Integration with ticketing or CRM systems for automated triage

---

Acknowledgements

This project was developed as part of the Portfolio Build at Purwadhika Digital Technology School — Data Science Bootcamp .

Special thanks to my teammate Ardinata Tambun for collaborating on this project during our first NLP modeling experience.