Supporting functionality to run 'caret' with spatial or spatial-temporal data. 'caret' is a frequently used package for model training and prediction using machine learning. This package includes functions to improve spatial-temporal modelling tasks using 'caret'. It prepares data for Leave-Location-Out and Leave-Time-Out cross-validation which are target-oriented validation strategies for spatial-temporal models. To decrease overfitting and improve model performances, the package implements a forward feature selection that selects suitable predictor variables in view to their contribution to the target-oriented performance. CAST further includes functionality to estimate the (spatial) area of applicability of prediction models by analysing the similarity between new data and training data.
Caret Applications for Spatio-Temporal models
This is the developer version of CAST. The CRAN Version can be found on https://github.com/environmentalinformatics-marburg/CAST
new feature: Best subset selection (bss) with target-oriented validation as (very slow but very reliable) alternative to ffs
minor adaptations: verbose option included, improved examples for ffs
bugfix: minor adaptations done for usage with plsr
new feature: Introduction to CAST is included as a vignette.
bugfix: minor error fixed in using user defined metrics for model selection.
ffs with option withinSE=TRUE did not choose a model as "best model" if it was within the SE of a model that was trained in an earlier run but had the same number of variables. This bug is fixed and if withinSE=TRUE ffs now only compares the performance to models that use less variables (e.g. if a model using 5 variables is better than a model using 4 variables but still in the SE of the 4-variable model, then the 4-variable model is rated as the better model).
plot_ffs plots the results of ffs to visualize how the performance changes according to model run and the number of variables being used.
Initial public version on CRAN