Chapter 11 Recommender Systems

A recommender system (also called a recommendation system or recommendation platform) is a subclass of information filtering systems that seeks to predict the rating or preference a user would give to an item.

Recommender systems are everywhere: in movie/series streaming platforms (e.g., Netflix), e-commerce websites (e.g., Amazon), social network platforms (e.g., Facebook, LinkedIn), or dating apps. These systems typically leverage historical data to automatically recommend items or products that a user is likely to be interested in.

So, let’s consider how those recommender systems are constructed. A useful starting point is the non-personalized baseline that many platforms still display: top-5 lists. A few years ago, the recommendations you could find were about top-5 destinations for the summer, top-5 Black Friday deals, and so on. Those lists are based on expert knowledge or aggregate statistics, and they lack personalization. This is exactly where recommender systems come in.

In this section, we explore fundamental recommendation methods: content-based methods, collaborative filtering (item-based or user-based), and latent factor models. If we classify the different methods, we notice that they are based on mainly two types of information: (i) characteristics or information about the items, and/or (ii) user-item interactions. This approach leads to

  • Content-based methods: focusing on attributes of a specific item.

  • Collaborative Filtering (item-based / user-based) methods: based on the analysis of user-item interaction data.

  • Baseline predictors: global, user, and item effects as a first model for the rating matrix.

  • Latent Factor Models: also based on the analysis of user-item interaction data.

Hybrid methods combine content-based ideas with collaborative filtering, and modern production systems often use deep learning. We mention those only briefly at the end; the rest of this chapter develops the three families above.

Several of these ideas were sharpened in industry, notably around the Netflix Prize (2006–2009), a public competition that asked teams to predict movie ratings and scored them by RMSE. Matrix factorization became a standard latent-factor method in large part because of that competition.