Context-Aware Recommendation

The same person wants different things on a Tuesday morning commute than on a Saturday night at home with friends. A recommender that knows only who you are and what exists treats those two requests identically. Context-aware recommendation refuses to.

From two dimensions to three
Classical recommendation predicts a score from a user and an item. Context-aware recommendation adds the situation of the request — time of day, day of week, location, device, weather, who the user is with — as a third input. Three families implement it. Contextual pre-filtering narrows the interaction data to the records matching the current context and runs an ordinary model on that slice. Contextual post-filtering ranks without context and then adjusts the list. Contextual modelling puts the context variables inside the model itself, as extra fields in a factorisation machine or a third mode in a tensor factorisation.

What changes in practice

• Short clips on a phone at eight in the morning and a feature film on a television at nine at night are the same user and genuinely different inventory.
• In food delivery, the ranking that works at lunch is wrong at midnight, and one global model averages the two into something that suits neither.
• Context makes data sparser, because every interaction now carries one more field — which is the trade-off the three families differ on.

Why it matters
Without it, preference that shifts by situation arrives at the model as unexplained noise and gets averaged away. The result is a recommender that is predictably wrong at predictable moments — holiday films in February, commute-length clips on a lazy Sunday — and whose offline metrics never reveal why.