tabnet is an R implementation of the TabNet deep learning architecture for tabular data, built on the torch package. It supports binary classification, multi-class classification, and regression tasks, with a tidymodels-compatible interface that accepts data frames, formulas, and recipes.
TabNet provides built-in interpretability through attention maps that show which features the model focuses on, viewable in aggregate or per network step. It handles missing predictor values natively through its masking mechanism, supports self-supervised pretraining for semi-labeled datasets, and includes an AUM loss function for imbalanced binary classification. It also supports hierarchical multi-label classification via coherent hierarchical networks, a capability not found in other TabNet implementations.
