Amazon Product Recommender (MIT IDSS – Capstone Project)

As part of a machine learning capstone from MIT-IDSS, this project developed a product recommendation system using a real-world dataset of Amazon product reviews. The goal was to design a recommender that improves customer engagement by suggesting relevant items based on user behaviour.

The work began with cleaning and preprocessing the dataset — removing duplicates, handling missing values, and addressing sparse user–item interactions. Through exploratory analysis, cold-start issues were investigated and user and product activity distributions were examined, establishing minimum rating thresholds to reduce noise from infrequent interactions.

Multiple models were implemented and evaluated, including:

• Rank-based recommender based on product popularity.
• User-based and item-based collaborative filtering with tuned similarity metrics.
• Matrix factorisation using SVD.
• A hybrid approach combining strengths of different models.

All models were evaluated using precision@k and recall@k. The final solution was an on-demand, deployable recommender using precomputed predictions for fast Streamlit inference.

Tech used: Python, Google-Colab, Streamlit | Surprise and Pickle (main libraries).

Project Links

Notebook Preview

Due to MIT-IDSS copyright restrictions, the notebook cannot be shared publicly. I am happy to discuss further details upon request.