E-commerce Recommendation System (Python, Amazon & Home Depot datasets)

Open Recommendation System - Paul.ipynb for the full analysis and code.

A well-designed recommendation system helps businesses improve the shopping experience on their site, which in turn improves customer acquisition and retention. This project builds one in three parts, following a new customer's actual journey: from their first visit with no purchase history, through their first purchase, to repeat visits once ratings and purchase history exist.

Datasets

1. Amazon product ratings from multiple users: Kaggle: Amazon Ratings
2. Home Depot product descriptions: Kaggle: Home Depot Product Search Relevance

Strategy

1. Popularity-based recommendations for new customers

A new customer with no purchase history is shown the most popular products on the site. This is a simple, effective way to cold-start a recommendation engine before any personal signal exists.

2. Model-based collaborative filtering

Once a customer has purchase history, recommendations switch to items based on that history and the ratings of other users who bought similar items. A model-based collaborative filtering approach was chosen here because it predicts products for a specific user by learning patterns across many users' preferences at once.

3. Item-item collaborative filtering (cold start for a new business)

For a business with no user-item purchase history at all, recommendations can instead come from a search-engine-style system based on textual clustering of product descriptions: predictions are drawn from other products in the same cluster for a given set of search terms.

---

Part I: Popularity-based recommendations for new customers

Analysis: The chart above ranks products by sales, descending. For example, product ID B001MA0QY2 has sales of over 7,000, followed by B0009V1YR8 at around 3,000, and so on.

Part II: Model-based collaborative filtering

Recommends the top 10 highly correlated products, in sequence, based on a customer's purchase history and the ratings of other customers who bought similar items.

Result: the products shown are the top 10 the recommendation system would surface to this customer, based on the purchase history of other customers on the site.

Part III: Item-item collaborative filtering (cold start for a new business)

For a business with no purchase or rating history yet, recommendations can be built from textual clustering of product descriptions instead.

Product clusters (subset of data):

Top terms per cluster:

Cluster 1	Cluster 2	Cluster 3
light	painted	water
watt	wood	concrete
volt	insulation	use
led	ft	ft
power	primed	metal
fan	proposition	used
bulb	nbsp	provides
bulbs	residents	designed
lighting	california	high
home		fiberglass

Once a cluster is identified from a customer's search terms, the recommendation system displays items from the matching product cluster based on product descriptions.

Summary

This approach works best when a business is setting up an e-commerce site for the first time and has no user-item purchase or rating history yet: it gives new visitors a reasonable starting recommendation. Once real purchase history accumulates, the system switches to model-based collaborative filtering for more personalized results.

License

No license file is currently attached to this repository; all rights reserved by default. Open an issue if you'd like to use this code and need clarification on terms.