User Bias
Two people give the same film four stars and mean entirely different things by it. One hands out fours freely; the other reserves anything above three for work they genuinely admire. User bias is the single number a factorisation model learns to absorb that habit, so the rest of the model can get on with taste.
Where it sits in the prediction
A biased factorisation model predicts a rating as the global mean, plus a user bias, plus an item bias, plus the dot product of the user and item factors. The user term is one scalar per user, learned by the same SGD or ALS pass that fits the factors and pulled toward zero by L2 regularisation — so a user with four ratings gets a cautious offset while a heavy rater gets the full measured one.
What it buys you
- The latent factors stop spending capacity encoding a constant offset and model preference instead.
- A user with almost no history still gets a sensible prediction from the three bias terms alone — on MovieLens and the Netflix Prize data, that baseline beats most naive methods outright.
- Scores become comparable across users, which matters when a threshold decides what gets sent or shown.
The ranking caveat
A user bias is constant across every item, so it cannot change the order of one user's recommendation list. It improves rating prediction and score calibration; it does nothing for top-N ranking from implicit feedback, where the item bias is the term doing the reordering.
User bias is a per-user offset in a factorisation model — one learned scalar capturing how high or low that user scores relative to the global mean, added alongside the item bias and the user-item dot product. It absorbs systematic rating style so the latent factors model genuine preference rather than habit, and regularisation keeps it small for users with little history. It improves rating accuracy; being constant across items, it cannot reorder a user's ranked list.
Some people are easy graders. Give them any decent film and they will call it four out of five. Others hand out a five once a year. Both might enjoy the same film equally, yet their scores look nothing alike.
User bias is how recommendation software allows for that. It learns one small number per person — roughly, this one is a generous rater, that one is hard to please — and takes it off the top before comparing anyone's taste with anyone else's. What is left is the part that actually says something about what you enjoy, rather than how freely you hand out stars.