23  Recommender Systems

Shuai Zhang (Amazon), Aston Zhang (Amazon), and Yi Tay (Google)

A catalog may contain millions of items, while any one user will inspect only a small fraction of them. A recommender system uses the interactions that have been observed—ratings, clicks, purchases, or viewing histories—to score or rank the remaining candidates for that user. Unlike search, which begins with an explicit query, recommendation must infer a useful ranking from partial and selectively observed behavior.

This setting creates two recurring difficulties. First, an unobserved user–item pair is not an observed dislike: the user may never have encountered the item. Second, the evaluation protocol determines the question being asked. A random interaction split measures a warm-start interpolation problem, whereas a chronological split asks whether past behavior predicts a later choice.

The chapter develops three families of tasks. Rating models predict explicit scores; ranking models learn from implicit interactions; and feature-rich models predict events such as clicks from user, item, and contextual fields. The progression from matrix factorization to AutoRec, NeuMF, Caser, factorization machines, and DeepFM shows how the input data and objective—not the presence of a neural network alone—determine what a model can learn.

The models can be compared by the evidence supplied to them and the question used for evaluation:

Model	Input	Training objective	Assumption about unobserved pairs	Evaluation used here
MF	Explicit ratings	Masked squared error	Omitted from the loss	RMSE on a random warm-start holdout
AutoRec	Partially observed rating vectors	Masked reconstruction error	Omitted from the loss	RMSE on the same holdout
NeuMF	User, observed item, sampled item	BPR pairwise loss	Sampled unobserved items act as comparisons	Hit rate and AUC on a chronological holdout
Caser	Ordered recent items and user ID	BPR pairwise loss	Same sampling assumption as NeuMF	Next-item ranking after chronological holdout
FM	Categorical impression fields	Binary log loss	Every row is an observed impression with a click label	Held-out log loss
DeepFM	The same impression fields	Binary log loss	The same labeled-impression assumption as FM	Held-out log loss