Product Review Rating System

A machine learning pipeline for automatically rating product reviews using RoBERTa, built with Kedro and MLflow.

Overview

This system processes product reviews and predicts rating scores (1-5) using a fine-tuned RoBERTa model. It includes data validation, model training, monitoring, and deployment capabilities.

Comprehensive Documentation

Quarto docs available from docs_quarto/_site/index.html

Features

• Data Processing: Handles CSV review data with text cleaning and preprocessing
• Model: Fine-tuned RoBERTa model for 5-class classification
• MLflow Integration: Experiment tracking and model versioning
• Data Monitoring: Drift detection and data quality validation using Evidently
• API Server: FastAPI-based prediction endpoint with logging
• Testing: Nodes tested using pytest which run with every build
• Docker Support: Containerized model serving

Main Project Structure

src/review_rating/pipelines
                    ├── data_processing/
                    │   ├── dataset.py       # PyTorch dataset implementation
                    │   └── nodes.py         # Data processing pipeline nodes
                    │   └── pipeline.py
                    ├── data_science/
                    │   └── nodes.py         # Model training and evaluation
                    │   └── pipeline.py
                    ├── data_monitoring
                    │   └── nodes.py
                    │   └── pipeline.py

Requirements

• Python 3.12
• PyTorch
• Transformers
• Kedro
• MLflow
• FastAPI
• Docker
• Quarto
• Pytest

Quick Start

1. Install dependencies:

pip install -r requirements.txt

2. Start the prediction server:

docker build -t review-rating-server .
docker run -d -p 5002:5002 --name review-rating-server review-rating-server

3. Run the pipeline:

kedro run

API Endpoints

• POST /predict: Get rating prediction for a review
• GET /batch-predict: Get rating predictions for a list of reviews
• GET /health: Chech if the server is ready
• GET /logs: Query model prediction logs

Model Details

• Architecture: RoBERTa-base
• Task: Multi-class classification (5 classes)
• Input: Review text (max length: 512 tokens)
• Output: Rating prediction (1-5)

Monitoring

The system includes data monitoring capabilities:

• Data drift detection
• Missing value analysis
• Duplicate detection
• Data quality validation