Ridge-regularized linear regression augments the least-squares criterion with a penalty on the magnitude of the coefficient vector, shrinking the estimated parameters toward zero and trading a small increase in training error for reduced sensitivity to noise. Because the penalty is applied to raw coefficient magnitudes, the procedure is not invariant to the scale of the input variables, so features must be normalized before regularization is meaningful; likewise unordered categorical attributes must first be converted to indicator variables via one-hot (nominal-to-binary) encoding before entering a linear model. Model quality is assessed by cross-validation rather than training-set reuse, and the regularization strength itself is chosen by a wrapping search over candidate values evaluated under that same cross-validation protocol. The material sits in applied supervised learning, specifically the workflow of fitting, regularizing, and validating linear models.
Ridge-regularized linear regression augments the least-squares criterion with a penalty on the magnitude of the coefficient vector, shrinking the estimated parameters toward zero and trading a small …