New A-spline method selects knots automatically with fewer basis elements
Matches P-spline performance while using a fraction of the basis functions...
A new preprint from Carrizosa, Guerrero, and Durbán tackles a long-standing problem in nonparametric modeling: automatically selecting the number and placement of knots for B-spline regression. Their method extends adaptive splines (A-splines) into generalized additive models (GAMs) by coupling knot selection with a specialized Fellner-Schall tuning algorithm for parameter optimization. The approach directly selects a sparse set of knots, unlike the regularization methods (e.g., P-splines) that dominate GAM practice.
In experiments across synthetic and real-world datasets, the technique matches the prediction performance of P-splines and leading knot-selection algorithms, while constructing models with significantly fewer basis elements—meaning leaner, more interpretable models with less risk of overfitting. The paper also notes that explicit knot selection can offer advantages in contexts where model transparency and parsimony are critical, such as in scientific or policy-driven applications. While the method introduces some computational overhead from the selection process, the reduction in basis complexity may offset this in large-scale settings.
- Authors propose a novel explicit knot selection technique for GAMs, extending adaptive splines (A-splines) with a customized Fellner-Schall tuning scheme.
- Achieves comparable prediction performance to P-splines and other state-of-the-art knot-selection methods across multiple synthetic and real datasets.
- Resulting models use a substantially smaller number of basis elements, improving interpretability and reducing overfitting risk.
Why It Matters
Simpler, more interpretable GAMs with automatic knot selection mean faster training and better model transparency for data scientists.