Guide to Building Online Recommendation Systems

“Candidate generation is responsible for quickly and cheaply narrowing the set of possible candidates down to a small enough set that can be ranked”

“At one or more points, the recommendation system will need to look up or compute data/features for the user and the candidates being considered. This data will fall into three categories: Item features, User features, User-Item cross fea…”

“Filtering is the process of removing candidates based on fetched data or model predictions. Filters primarily act as system guardrails for bad user experience”

“online model inference boils down to sending feature vectors to a model service to get predictions. Typically these models predict easily measurable downstream events, like Click Through Rate, Probability of a Purchase, Video Watch Time”

“Pointwise ranking is the process of scoring and ranking items in isolation; i.e., without considering other items in the output. The item pointwise score may be as simple as a single model prediction (e.g., the predicted click through ra…”

“Listwise ranking is the process of ordering items in the context of other items in the list. ML-based and heuristic-based approaches can both be very effective for listwise optimization”