virheenkäyrän

Virheenkäyrä, literally error curve in Finnish, is a graphical representation used in statistical and machine learning contexts to illustrate the relationship between predictive error and a chosen model parameter or complexity measure. It typically plots a measure of prediction error, such as mean squared error, mean absolute error, or classification accuracy, on the vertical axis against a parameter value on the horizontal axis. Commonly examined parameters include regularization strength, number of hidden units, depth of a decision tree, or smoothing bandwidth.

The basic idea behind a virheenkäyrä is to help diagnose underfitting and overfitting. When error is high at both low and high parameter values but lower in the middle, the curve exhibits a U‑shaped pattern that indicates an optimal parameter choice near the curve’s minimum. If error steadily decreases without plateaus as complexity increases, it may suggest that more data or additional regularization is needed.

In experimental design and model selection, researchers often employ cross‑validation to estimate error values for the curve. By averaging over folds, the virheenkäyrä reflects expected performance on unseen data rather than solely on a training set. It is additionally used to compare models: overlapping curves may show similar predictive trade-offs, while diverging curves highlight distinct behaviors.

The term appears frequently in Finnish scientific literature on predictive modeling, with the objective of aiding transparent decision‑making regarding model hyperparameters. It aligns with analogous concepts in English‑speaking literature such as learning curves, validation curves, and regularization paths.