Fig. 2. Workflow for domain of applicability (DA) identification and validation for an ML model.
The DA is described by a selector (σf) that is comprised of logical conjunctions of a representation space (here symbolized by a single dimension x for simplicity but may be multidimensional). The selector is identified by applying subgroup discovery (SGD) to the individual ML-model errors for subset of test set (DA identification set). An unbiased estimate of the model performance within the DA is obtained on the remaining samples of the test set that were left out of the DA identification (DA validation set).