How does recommendation software work (like Netflix or Spotify)?

The goal of recommendation systems

A recommendation system (or recommendation engine) is an AI system that predicts which content, products, or people are relevant to a specific user. The goal: keep the user on the platform longer and increase the chance of a purchase or interaction.

Illustration created with Canva AI

Two main approaches

Collaborative filtering

Users who are similar to you — comparable viewing history, musical taste — have also liked other things. The system recommends those items based on the similarities between users.

Simply put: "users who watched this also watched that".

Content-based filtering

The system analyzes the characteristics of items you liked and looks for items with similar properties. If you watched three science documentaries, it recommends more science documentaries.

Matrix factorization and deep learning

Modern systems like those of Netflix and Spotify go further. They use matrix factorization to discover latent (hidden) properties of users and items, and combine this with deep learning models that process contextual signals: time of day, device, session duration, what you just watched.

The filter bubble problem

Recommendation systems are designed to maximize relevance, not diversity. This leads to filter bubbles: you are increasingly exposed to content that confirms what you already believe or like, and less to surprising or challenging content.

Researchers and regulators are looking more critically at the societal effects of this mechanism, particularly on news platforms and social media.

Applications

• Netflix (70% of viewing time comes from recommendations)
• Spotify Discover Weekly
• Amazon product recommendations
• TikTok For You page
• LinkedIn job recommendations