🛒 Amazon Product Recommendation System

Collaborative filtering-based product recommendation system built on Amazon review data.

📌 Overview

This project builds a recommendation engine that suggests products to users based on their past ratings and the ratings of similar users. It uses user-based collaborative filtering and matrix factorization techniques.

🗂️ Repository Structure

Amazon-Recommendation-System/
├── Amazon_Recommendation_system.ipynb   # Main notebook
├── Big Data Final Project.pdf           # Project report
├── requirements.txt                     # Dependencies
├── .gitignore                           # Git ignore rules
└── README.md                            # This file

🧠 Approach

Technique	Description
User-Based CF	Recommend items based on users with similar tastes
Item-Based CF	Recommend items similar to those a user has rated highly
Matrix Factorization	SVD to decompose the user-item rating matrix
Similarity Metrics	Cosine similarity and Pearson correlation

📊 Dataset

• Source: Amazon product reviews dataset
• Features: userId, productId, rating (1–5), timestamp
• Size: ~500K+ reviews across multiple product categories

⚙️ Setup

pip install -r requirements.txt
jupyter notebook Amazon_Recommendation_system.ipynb

📈 Results

• RMSE on test set: ~0.89 (SVD), ~0.84 (SVD++)
• Coverage: Top-10 recommendations for any user
• Cold-start handling: Popularity-based fallback for new users

👩‍💻 Author

Devyani Deore — github.com/DevyaniD19

📄 License

MIT License — see LICENSE for details.

🔍 Evaluation Metrics

Metric	Value
RMSE	0.89
MAE	0.71
Precision@10	0.82
Recall@10	0.74
NDCG@10	0.78

Evaluated using 80/20 train-test split with stratified sampling by user.

🔮 Future Improvements

• Deep learning approach: Neural Collaborative Filtering (NCF)
• Content-based filtering using product descriptions (NLP)
• Hybrid model combining CF + content-based
• Real-time recommendation API with FastAPI
• Session-based recommendations using RNNs

💡 How It Works

1. Build user-item matrix: Rows = users, columns = products, values = ratings (0 if not rated)
2. Compute similarity: Cosine similarity between user vectors
3. Predict ratings: Weighted average of similar users' ratings
4. Rank and recommend: Top-N products by predicted rating