A retailer's return model was trained when most orders were in-store pickup. Ship-to-home orders have since become the common case, so the feature mix being scored no longer matches training, while the link from a given basket to a return has not been shown to change. What kind of drift is described?

Data drift means the distribution of the features has changed relative to the data the model was trained on. Pickup used to dominate, and ship-to-home is now the common case, so the mix of inputs being scored has moved. The stem also says the link from a given basket to a return has not been shown to change. The relationship the model learned can still be the right relationship, applied to a different mix of baskets. That is data drift, not a new concept. Concept drift would be a change in that feature-to-target relationship. The option that says any change in channel frequency rewrites the basket-to-return link ignores the stem's statement that the link was not shown to move. A shift in mix is not, by definition, a new mapping. Label leakage would mean the target, or a stand-in for the target, was copied into the inputs. A change in how orders are fulfilled does not require the return flag to be hidden inside the features. A cold start is an empty history for a new user or item in a recommender. A more common channel is a change in the feature distribution of orders the model already knows how to describe, not an item with no interactions. Exam caveat: if a later study shows that the same basket now returns at a different rate, that additional finding would be concept drift. It is not what this stem establishes. Operational check: compare training and recent distributions of the channel feature, and separately check return rates within similar baskets before you decide which drift you have.