Building Recommendation Systems in Python and JAX

Chapter 12. Training for Ranking

Typical ML tasks usually predict a single outcome, such as the probability of being in a positive class for classification tasks, or an expected value for regression tasks. Ranking, on the other hand, provides a relative ordering of sets of items. This kind of task is typical of search results or recommendations, where the order of items presented is important. In these kinds of problems, the score of an item usually isn’t shown to the user directly but rather is presented—maybe implicitly—with the ordinal rank of the item: the item at the top of the list is numbered lower than the next item.

This chapter presents various kinds of loss functions that ML algorithms can use during training. These scores should estimate list orderings such that when compared to one another, they result in sets that are ordered more closely to the relevance ordering observed in a training dataset. Here we will focus on introducing the concepts and computations, which you’ll put to work in the next chapter.

Where Does Ranking Fit in Recommender Systems?

Before we dive into the details of loss functions for ranking, we should talk about where ranking fits into the larger scheme of recommender systems as a whole. Typical large-scale recommenders have a retrieval phase, in which a cheap function is used to gather a decent number of candidate items into a candidate set. Usually, this retrieval phase is only item based. For example, the candidate set might include items ...