Item Bias
Some films are simply better liked than others, by nearly everyone. That has nothing to do with personal taste, and a model that tries to explain it with taste wastes its capacity doing so. The item bias is the single number that takes the job off its hands.
A per-item offset, learned like everything else
In a biased factorization model a prediction is the global mean plus a user bias plus an item bias plus the dot product of the user and item factors. The item bias carries how far this item sits above or below the overall average — a well-made documentary at +0.7 stars, a forgettable sequel at −0.4 — and it is learned by the same optimisation as the factors rather than computed in advance as an average.
Why fitting it beats averaging it
A raw mean rating is unreliable for an item with four ratings, and misleading if those four came from unusually generous users. Fitting the bias jointly, with L2 regularisation pulling it toward zero, shrinks thinly observed items back toward the average in proportion to how little evidence supports them. This is why mean-plus-biases is a strong baseline on MovieLens, and the one a factorization model must beat to show it learned preference.
What it leaves to the factors
With the overall level absorbed, the factors are free to model disagreement — who likes this and who does not — which is where personalisation actually lives. Drop the bias term and crowd-pleasers drift up every user's list, and the model confuses general quality with individual taste.
Item bias is a learned per-item offset in a factorization model, capturing how far an item's reception sits above or below the global mean independently of who is rating it. Predicting the mean plus user and item biases plus the factor interaction lets the factors model disagreement rather than overall quality. Fitted jointly under L2 regularisation, it shrinks thinly rated items toward the average instead of trusting a noisy observed mean.
Some films are loved by nearly everyone and some are widely disliked, whatever your personal taste happens to be. A recommender needs a way of saying that plainly, instead of pretending it is about you.
Item bias is that number: a small adjustment attached to each item, recording how much better or worse it goes down with people in general. Once the system has it, the rest of the machinery is free to concentrate on the interesting part — where people disagree. Without it, crowd-pleasers would quietly float to the top of everybody's list, and the system would mistake general popularity for something it had learned about you.