Item Neighborhood

Ask why a shop suggested a particular paperback and the honest answer is usually "because of the other books you bought". The set of books doing that work is the item neighborhood.

What the neighborhood is
For a candidate item, its neighborhood is the set of catalogue items most similar to it under a stored similarity measure — adjusted cosine on ratings, cosine or conditional probability on implicit counts. The full similarity matrix is never kept: each item retains only its top k most similar items, and everything outside that list is treated as similarity zero. That truncated structure is what gets precomputed offline and loaded into the serving layer.

How a score comes out of it
To score an item for a user, the system intersects the item's neighborhood with the items that user has already touched, then takes a similarity-weighted average of the user's ratings over that intersection. No user-level parameters are involved anywhere, so an interaction from one minute ago influences the next request without any retraining.

Why its size decides everything
The neighborhood is the entire capacity of an item-based recommender, and k is the dial that sets it:

• Too wide, and weakly related popular items leak in and flatten every list toward the bestseller shelf.
• Too narrow, and the intersection with a short user history is frequently empty, so no prediction can be produced at all.
• Sparsity, not taste, usually decides the outcome — where most items carry a handful of interactions, neighborhoods stay unreliable however k is tuned, and a learned latent factor model becomes the better answer.