RitSpls: A Julia package for Robustness-Inducing Transformations in SPLS regression

RitSpls implements a robustification of Sparse Partial Least Squares (SPLS) through robustness-inducing transformations on the (univariate) predictand and predictors.

The approach considered in [1] is:

1. transform multivariate predictors through a Generalized Spatial Sign Pre-Processing (GSSPP) [2];

2. transform the univariate predictand through a Wrapping transformation [3];

3. use the SNIPLS (sparse NIPALS) algorithm [4] to estimate an SPLS model from the (robustly) transformed inputs.

The package provides an SPLS class that allows to interface with the ScikitLearn API. For instance, objects of the SPLS and GSSPP class accept widely used ScikitLearn functions and routines (e.g, fit!, predict, transform, GridSearchCV, etc.).

]add <path to this GitHub repo>

Examples on how to use RitSplsare presented as Jupyter notebooks in the documentation folder.

[1] Serneels, S., L. Insolia, and T. Verdonck (2023). “Elegant robustfication of sparse partial least squares by robustness-inducing transformations". Submitted.

[2] Raymaekers, J. and P. J. Rousseeuw (2019). “A generalized spatial sign covariance matrix”. In: Journal of Multivariate Analysis 171, pp. 94–111.

[3] Raymaekers, J. and P. J. Rousseeuw (2021). “Fast robust correlation for high-dimensional data”. In: Technometrics 63 (2), pp. 184–198.

[4] Hoffmann, I., P. Filzmoser, S. Serneels, and K. Varmuza (2016). “Sparse and robust PLS for binary classification”. In: Journal of Chemometrics 30.4, pp. 153–162.

Functions for robustness-inducing transformations are adapted from R code developed by ‪Jakob Raymaekers‬.