Polynomial, Spline and Generalized Additive Models

Description

Models curved relationships, such as satisfaction rising then falling with feature count, instead of forcing straight lines.

Many UX relationships are not linear: ease ratings may rise with experience then plateau, conversion may peak at a moderate number of options, page speed may matter only past a threshold. Polynomial terms add curvature, splines fit flexible curves joined at knots, and generalised additive models (GAMs) learn a smooth curve for each predictor automatically while staying interpretable. Bruce and Bruce show how partial residual plots reveal the need for these models.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Plot each predictor against the outcome or partial residuals to spot curvature.
  2. Try polynomial terms for simple curves, or splines with a few knots.
  3. Fit a GAM to learn smooth functions per predictor.
  4. Compare models with cross-validation to avoid overfitting.
  5. Plot the fitted curves and describe thresholds or plateaus in plain language.

Sources

  • Practical Statistics for Data Scientists (Bruce & Bruce)

Also known as

  • GAMs
  • Spline regression
  • Non-linear regression
  • LOESS
  • LOWESS
  • Local regression
  • Kernel regression
  • Smoothing spline

Categories

  • Analytics & Behaviour
  • Statistics & Modelling