Amazon Product Review Analysis

This repository contains an end-to-end Amazon Product Review Analysis project, demonstrating how to build and deploy LLM-powered sentiment analysis using both BERT (from Hugging Face Transformers) and a custom LSTM. The project also includes steps to design an ad ranking system leveraging sentiment scores, and guidance for AWS cloud deployment using Lambda or SageMaker.

---

Table of Contents

1. Project Overview
2. Dataset & Preprocessing
3. Modeling Approaches
   • BERT (Transformers)
   • Custom LSTM
   • Three-Class Classification vs. Regression
4. Ad Ranking System
5. Deployment
   • AWS SageMaker
   • AWS Lambda
6. Visualizations
7. Repository Structure
8. How to Run
9. Future Work
10. License

---

Project Overview

Goal: Develop a comprehensive pipeline that performs:

• Sentiment Analysis of Amazon product reviews (negative, neutral, positive), or direct star rating regression (1–5).
• Ad Ranking System using sentiment scores to improve user engagement and click-through rates (CTR).
• Cloud Deployment for scalable, low-latency inference on AWS (Lambda or SageMaker).

This project was inspired by real-world scenarios where companies like Amazon, Apple, Netflix, etc., use machine learning to optimize user experience and product recommendations.

---

Dataset & Preprocessing

1. Dataset: The project uses a .jsonl.gz file of Amazon product reviews (e.g., Movies_and_TV.jsonl.gz), which contains fields like:

   • rating (1–5 star rating)
   • text (the review content)
   • Other optional metadata (e.g., title, timestamp, asin, etc.)

2. Preprocessing:

   • Parsing: We read the GZ file line-by-line using Python’s gzip and json.
   • Cleaning: Convert text to lowercase, remove punctuation, etc. (especially for LSTM tokenization).
   • Labeling:
     • 3-Class Classification:
       • 0 = Negative (rating ≤ 2),
       • 1 = Neutral (rating = 3),
       • 2 = Positive (rating ≥ 4).
     • Regression: Predict the exact star rating from 1 to 5.

3. Train-Test Split: We split the dataset into 80% training and 20% test (or other ratios) using sklearn.model_selection.train_test_split, with stratification by sentiment if doing classification.

---

Modeling Approaches

BERT (Transformers)

• Library: Hugging Face Transformers (transformers package).
• Model: BertForSequenceClassification for 3-class classification or regression (num_labels=1 with problem_type="regression").
• Tokenization: BertTokenizer (bert-base-uncased by default).
• Training: Fine-tuned with AdamW optimizer, typical learning rate of 2e-5 for a few epochs.
• Saving: model.save_pretrained() + tokenizer.save_pretrained().

Custom LSTM

• Architecture: An embedding layer -> LSTM -> linear layer for output.
• Preprocessing: Build a vocabulary (word_to_idx), convert each review to sequences of word indices, handle padding.
• Training: Use CrossEntropyLoss for classification (3-class) or MSELoss for rating regression.
• Saving: Use torch.save(model.state_dict(), "lstm.pth").

Three-Class Classification vs. Regression

1. 3-Class:
   • Negative, Neutral, Positive
   • Helpful if you want discrete categories for your ad ranking logic.
2. Regression:
   • Predict star rating from 1.0 to 5.0 (continuous).
   • Offers finer granularity; can rank products by exact predicted rating.

---

Ad Ranking System

1. Predict Sentiment or Rating: For each product review, use the trained model to generate a sentiment score or predicted star rating.
2. Aggregate Scores:
   • For each product (identified by asin), compute the average predicted sentiment/ rating across all its reviews.
3. Ranking:
   • Sort products in descending order of average sentiment/rating.
   • (Optional) Combine with other signals like historical CTR or user context.
4. CTR Improvement:
   • Hypothetically, we measure CTR via an A/B test comparing the new ranking method to a baseline. In many real-world scenarios, a 15% improvement can be achieved with better product targeting.

---

Deployment

AWS SageMaker

1. Save Model:

   model.save_pretrained('bert_3class_model')
   tokenizer.save_pretrained('bert_3class_tokenizer')

2. Upload the artifacts to S3 using awscli or boto3.
3. Create SageMaker Endpoint:
   • Define a PyTorchModel or HuggingFaceModel referencing your S3 artifacts.
   • Provide an inference.py script that loads the model and handles prediction.
   • Deploy to an instance type like ml.m5.large.

AWS Lambda

1. Package a smaller model (e.g., DistilBERT or a minimal LSTM) into a ZIP under 250MB (unzipped).
2. Upload to AWS Lambda, potentially using a Lambda Layer for large dependencies (like torch).
3. API Gateway can trigger the function for on-demand sentiment predictions.