2 Anatomy of a recommender system

When thinking about recommender systems, it’s often helpful to consider them as a single operation: given a user’s request for recommendations—for example, when opening a streaming app—they provide a ranked list of suggestions relevant to the user in question. This is called the black box model, a useful abstraction when the details are not important and we want to focus only on the input and output. To better illustrate how things work, let’s now peek under the hood at their main components. This exploration will also uncover the various places where decisions are made about the content we see.

2.1 The user-and-item world

2.2 Parts of a RecSys

2.2.1 Candidate generation

2.2.2 Filtering

2.2.3 Heavy ranking

2.2.4 Post processing

2.2.5 Delivering and recording

2.3 Heavy ranking revisited: Training a RecSys

2.3.1 Heuristics

2.3.2 Data-driven approaches

2.4 How to train a machine learning model

2.5 Raw and derived scores

2.6 Platforms’ design and amplification

2.7 What happens when you refresh your feed

2.8 Summary