dynamic feature selection, where we sequentially query features to make
accurate predictions with a minimal budget, is a promising paradigm to reduce
feature acquisition costs and provide transparency into the prediction process.
The problem is challenging, however, as it requires both