Learning Paradigms and Problem Framing

Modelling the joint gives you a distribution over the features themselves, so you can draw synthetic records, integrate out a feature missing at scoring time, and judge how unusual an input is. A p(y|x) model answers only the labelling question.

Domain experts write cheap rules that each vote a class on a row or abstain. The votes are combined, weighted by each rule's estimated accuracy, into probabilistic labels, and a model trained on them generalises beyond the rules.

A lazy model does its work at request time against everything it has stored, so memory and per-query cost both scale with the store. Bound what is stored, or distil the same data into a fixed-size eager model.

Freeze the model when validating a new version costs more than staleness does, when a past decision must be explainable against a specific artifact, or when its outputs shape future labels. Freezing trades silent decay for reviewable behaviour.

Almost always one global model. Pooling 240,000 store-SKU series into a single training table, with store and product identity as features, shares patterns across short and sparse histories, handles new products, and leaves one artifact to operate instead of 240,000.

Ask whether the action changes the situation the next decision faces. If each impression pays its own reward and the next starts from the same state, it is a contextual bandit; only real carry-over justifies reinforcement learning.