Amazon product recommendation system
I built and compared several ways to recommend electronics to Amazon customers, using a dataset of product ratings.
The problem
Shoppers face too many choices. A recommendation system needs to predict how a customer would rate a product they haven’t tried yet.
The approach
I built a rank-based model, user-user and item-item collaborative filtering, and matrix factorization. I then tuned their hyperparameters and compared them on precision, recall, and F1 score.
The result
Tuned matrix factorization had the highest F1 score, at 0.829, a small gain over the 0.827 before tuning.
What I’d change
I’d tune the hyperparameters further and try combining these methods with other recommendation techniques.