Contextual Post-Filtering
Run the recommender you already have, ignore the situation entirely, take the ranked list it produces — then fix the list up before anyone sees it. That is contextual post-filtering, and its appeal is that the model underneath never has to change.
Adjust the output, not the model
A conventional context-free model scores the candidate set. The current context is then applied to that output in one of two ways. Filtering removes items unsuitable for the context outright. Weighting multiplies each score by how relevant the item's attributes are in this context, usually estimated from how often items with those attributes were engaged with in that context historically. Either way the ordering of a context-free model is the starting point and context only acts on it.
Where it fits
- Hard constraints at the end of a pipeline: out-of-stock items, adult titles on a shared family device, three-hour films at eleven at night. The re-ranking stage of a multi-stage recommender is exactly where this lives.
- It is cheap to ship — no retraining, no new model features, and a rollback that is one flag.
- It composes with anything, including a contextual model, because it touches only the output.
What it cannot do
It can only reorder or remove what retrieval already surfaced. If the right item for this situation never entered the candidate set, no post-filter can rescue it, because context never influenced retrieval. Aggressive filtering also thins the list, which is why the strict variants are reserved for genuine constraints and the soft ones for score weighting.
Contextual post-filtering applies context after a context-free model has ranked the candidates, either removing items unsuitable for the current situation or re-weighting their scores by how relevant their attributes are in that context. It needs no retraining and composes with any underlying model, which makes it the usual home for hard constraints at the re-ranking stage. Its limit is that it can only reorder what retrieval already surfaced.
Think of a shop assistant who picks out ten jackets she thinks you would like, then glances out of the window, sees snow, and quietly puts the thin linen one back on the rail.
Contextual post-filtering works like that second glance. The software first chooses what to show you without worrying about the moment — the time, the place, the device in your hand — and only then removes or reshuffles the things that do not suit it. The advantage is that nothing underneath has to be rebuilt, so it is easy and safe to add. The limit is just as simple: it can only tidy up the shortlist it was handed, so anything left off that list stays off.