Structured Dimensionality Reduction for Additive Model Regression

摘要: Additive models are regression methods which model the response variable as sum of univariate transfer functions input variables. Key benefits additive their accuracy and interpretability on many real-world tasks. however not adapted to problems involving a large number (e.g., hundreds) variables, they prone overfitting in addition losing interpretability. In this paper, we introduce novel framework for applying The key idea is reduce task dimensionality by deriving small new covariates obtained linear combinations inputs, where weights estimated with regard problem at hand. moreover constrained prevent facilitate interpretation derived covariates. We establish identifiability proposed under mild assumptions present an efficient approximate learning algorithm. Experiments synthetic data demonstrate that our approach compares favorably baseline terms accuracy, while resulting lower complexity yielding practical insights into high-dimensional Our broadens applicability maintaining potential provide insights.