MF Regularization

A matrix factorisation fitted to a user with four ratings will happily invent a dramatic personality for them: given enough freedom the factors reproduce those four numbers exactly and predict nonsense everywhere else. Regularisation is the term in the objective that refuses to let that happen.

The mechanics
The fitted objective is squared error on the observed entries plus a penalty on the size of the parameters — a coefficient times the sum of squared norms of every user factor, item factor and bias. Minimising the sum means a factor value may only grow if the error it removes outweighs the penalty it adds. Entities with little data cannot clear that bar, so their factors shrink towards zero and their predictions fall back on the global mean and the bias terms. Shrinkage is not a side effect here; it is the mechanism.

Choices that matter in practice

Biases usually get their own coefficient, since a well-observed item bias deserves less shrinking than a speculative factor vector.
Scaling each entity's penalty by its number of observations — the weighted-lambda regularisation introduced in ALS-WR during the Netflix Prize — shrinks a four-rating user hard while leaving a four-thousand-rating user largely alone.
The coefficient is tuned on a validation split, and its optimum moves with rating density.

Why it is load-bearing here
Density in recommendation is around one percent and the item distribution is a long tail, so most parameters in the model rest on a handful of observations. Without the penalty, validation error diverges within a few epochs and tail items receive wild, confident scores that push them into top-N lists where they do not belong. With it, an unknown entity simply looks average instead of extreme.

MF regularisation adds an L2 penalty on user factors, item factors and bias terms to the squared-error objective of a matrix factorisation, scaled by a coefficient tuned on validation data. It shrinks parameters estimated from few observations towards zero, so sparsely-observed users and tail items fall back on the global and bias terms instead of receiving extreme scores. Without it, a model fitted at roughly one percent density overfits within a few epochs.

Think of a reviewer who visits a restaurant once and declares it the best in the city. A sensible editor tones that down — one visit is not enough to justify a claim that strong — while leaving the verdict of someone who has eaten there fifty times intact.

Regularisation is that editor, written into the maths. When a recommendation system learns a description of each person and each item, this rule stops it making bold claims from thin evidence. Someone who has rated four films gets a cautious, near-average profile; someone with four hundred ratings gets a detailed one. It is the difference between a system that guesses wildly about newcomers and one that admits it has nothing remarkable to say yet.