Amazon.com Recommendations: Item-to-Item Collaborative Filtering

Inside the paper

The problem

A large shop must generate useful recommendations quickly despite a large catalogue and many customers. Comparing a visitor with the entire customer base at request time becomes expensive as the business grows.

The idea

Amazon’s approach calculates relationships between items in advance and uses those relationships to assemble recommendations for a particular customer. Items bought or rated by the customer become starting points for finding related products.

How it works

• Build a table of related items offline by examining which products are purchased by the same customers.
• Calculate item similarity from customer purchase vectors; the paper describes cosine similarity as one way to do this.
• For an active customer, look up items similar to their purchases or ratings and combine the candidates into a recommendation list.
• The paper describes ways to reduce the offline cost, including sampling customers for very popular items, and discusses how input size affects online work.

What it showed

The authors describe an operational recommendation system used across Amazon’s site, including homepage and shopping-cart examples. They compare its scaling characteristics with other approaches, but do not present a controlled numerical comparison of recommendation quality in this report.

Read it for

Start with the item-to-item section on pages 3–4 and its pseudocode to see how the table is built. The earlier discussion of alternative methods and the later scaling analysis are useful context; the interface screenshots can be skimmed.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.