Amazon Review Sentiment Analysis – MLOps Pipeline

Overview

This project implements a complete MLOps pipeline for sentiment analysis on Amazon product reviews using Azure Machine Learning. The workflow includes data preprocessing, model training, hyperparameter tuning, evaluation, deployment, and inference via a REST API endpoint.

---

Model Choice

The model used is Logistic Regression (scikit-learn).

Rationale

• Efficient and fast to train on large datasets
• Suitable for binary classification tasks
• Provides a strong baseline for text-based features
• Simple to deploy and interpret

---

Feature Engineering

The model uses engineered numerical features derived from the dataset.

Key steps

• Expanded list-based features into separate columns

• Removed non-numeric and irrelevant columns such as:

  • reviewText, title, summary, asin, reviewerID

• Converted all features to numeric format

• Filled missing values with 0

• Ensured consistent feature columns across train, validation, test, and deployment datasets

Feature Sets Tested

• sbert_tfidf
• all (selected for final model)

---

Hyperparameter Tuning

Hyperparameter tuning was performed using an Azure ML sweep job.

Parameters tuned

• Regularization strength (C)
• Solver type
• Maximum iterations

Outcome

• Minimal variation across runs
• Performance remained consistent
• Final configuration selected based on stability

---

Model Performance

Training, Validation, and Test Results

• Accuracy: ~0.802
• Precision: ~0.802
• Recall: 1.0
• F1 Score: ~0.890
• AUC: ~0.50

Observations

• High recall indicates the model predicts most positive cases correctly
• AUC near 0.5 suggests limited discrimination capability
• Model tends to favor the majority class

---

Model Deployment

The trained model was deployed using Azure ML Managed Online Endpoints.

Deployment Components

• Registered model: amazon-review-sentiment-model
• Scoring script: score.py
• Environment: custom conda environment with required dependencies
• Compute: Standard_F2s_v2

Endpoint

• Exposes a REST API for real-time predictions
• Secured using API key authentication

---

Inference and Endpoint Invocation

A custom script (invoke_endpoint.py) was used to:

• Load the deployment dataset
• Apply the same feature engineering pipeline used during training
• Send HTTP POST requests to the endpoint
• Receive predictions
• Compute evaluation metrics

---

Deployment Performance

Results on unseen deployment data:

• Accuracy: 0.802
• Precision: 0.802
• Recall: 1.0
• F1 Score: 0.890

Interpretation

• Performance is consistent with the test dataset
• No significant degradation observed, indicating minimal data drift

---

MLOps Workflow Summary

This project includes:

• Data preprocessing and feature engineering
• Model training using Azure ML compute
• Hyperparameter tuning with sweep jobs
• Model evaluation (train, validation, test)
• Model registration
• Deployment via managed online endpoint
• Real-time inference using REST API
• Post-deployment evaluation

---

Limitations and Future Work

• The model shows bias toward the majority class

• AUC indicates limited predictive separation

• Potential improvements:

  • Enhanced feature engineering
  • More advanced models (e.g., transformer-based approaches)
  • Techniques to address class imbalance

---

Conclusion

This project demonstrates a complete end-to-end MLOps pipeline using Azure Machine Learning. The model is successfully trained, deployed, and used for real-time inference, fulfilling all assignment requirements.