An R implementation of: TabNet: Attentive Interpretable Tabular Learning (Sercan O. Arik, Tomas Pfister).
The code in this repository is an R port using the torch package of dreamquark-ai/tabnet PyTorch’s implementation.
TabNet is augmented with Coherent Hierarchical Multi-label Classification Networks (Eleonora Giunchiglia et Al.) for hierarchical outcomes.

Installation

You can install the released version from CRAN with:

The development version can be installed from GitHub with:

# install.packages("remotes")
remotes::install_github("mlverse/tabnet")

Basic Binary Classification Example

Here we show a binary classification example of the attrition dataset, using a recipe for dataset input specification.

library(tabnet)
suppressPackageStartupMessages(library(recipes))
library(yardstick)
library(ggplot2)
set.seed(1)

data("attrition", package = "modeldata")
test_idx <- sample.int(nrow(attrition), size = 0.2 * nrow(attrition))

train <- attrition[-test_idx,]
test <- attrition[test_idx,]

rec <- recipe(Attrition ~ ., data = train) %>% 
  step_normalize(all_numeric(), -all_outcomes())

fit <- tabnet_fit(rec, train, epochs = 30, valid_split=0.1, learn_rate = 5e-3)
autoplot(fit)

The plots gives you an immediate insight about model over-fitting, and if any, the available model checkpoints available before the over-fitting

Keep in mind that regression as well as multi-class classification are also available, and that you can specify dataset through data.frame and formula as well. You will find them in the package vignettes.

Model performance results

As the standard method predict() is used, you can rely on your usual metric functions for model performance results. Here we use {yardstick} :

metrics <- metric_set(accuracy, precision, recall)
cbind(test, predict(fit, test)) %>% 
  metrics(Attrition, estimate = .pred_class)
#> # A tibble: 3 × 3
#>   .metric   .estimator .estimate
#>   <chr>     <chr>          <dbl>
#> 1 accuracy  binary         0.837
#> 2 precision binary         0.837
#> 3 recall    binary         1
  
cbind(test, predict(fit, test, type = "prob")) %>% 
  roc_auc(Attrition, .pred_No)
#> # A tibble: 1 × 3
#>   .metric .estimator .estimate
#>   <chr>   <chr>          <dbl>
#> 1 roc_auc binary         0.546

Explain model on test-set with attention map

TabNet has intrinsic explainability feature through the visualization of attention map, either aggregated:

explain <- tabnet_explain(fit, test)
autoplot(explain)

or at each layer through the type = "steps" option:

autoplot(explain, type = "steps")

Self-supervised pretraining

For cases when a consistent part of your dataset has no outcome, TabNet offers a self-supervised training step allowing to model to capture predictors intrinsic features and predictors interactions, upfront the supervised task.

pretrain <- tabnet_pretrain(rec, train, epochs = 50, valid_split=0.1, learn_rate = 1e-2)
autoplot(pretrain)

The example here is a toy example as the train dataset does actually contain outcomes. The vignette on Self-supervised training and fine-tuning will gives you the complete correct workflow step-by-step.

Missing data in predictors

{tabnet} leverage the masking mechanism to deal with missing data, so you don’t have to remove the entries in your dataset with some missing values in the predictors variables.

Comparison with other implementations

Group Feature {tabnet} dreamquark-ai fast-tabnet
Input format data-frame
formula
recipe
Node
missings in predictor
Output format data-frame
workflow
ML Tasks self-supervised learning
classification (binary, multi-class)
regression
multi-outcome
hierarchical multi-label classif.
Model management from / to file v
resume from snapshot
training diagnostic
Interpretability
Performance 1 x 2 - 4 x
Code quality test coverage 85%
continuous integration 4 OS including GPU

Alternative TabNet implementation features