Table 2 Policy Evaluation: Mean Average Error (\(\mu\) ± \(\sigma\)) of self-normalized inverse propensity scoring (SNIPS)-based estimators. Optdigits and Letter are two multiclass classification datasets from the UCI repository [30]. LR Logistic Regression, BNN Bayesian Neural Network, NN Neural Network
From: Clinical decision making under uncertainty: a bootstrapped counterfactual inference approach
Dataset | Expert Policy | SNIPS(\(h^{true}_0\)) | \(\hat{h}_0\) - NN | \(\hat{h}_0\) - BNN | |||||
|---|---|---|---|---|---|---|---|---|---|
Vanilla SNIPS | NN Ensemble | Vanilla SNIPS | BNN (Variational Inf.) | MC-Dropout | |||||
SNIPS\(_{inv}\) | SNIPS\(_{avg}\) | SNIPS\(_{inv}\) | SNIPS\(_{avg}\) | SNIPS\(_{avg}\) | |||||
UCI | OPTDIGITS (10 actions) | 2.6 ± 0.4 | 38.7 ± 25.2 | 17.6 ± 6.7 | 1.0 ± 0.2 | 5.5 ± 2.8 | 11.2 ± 4.1 | 1.1 ± 0.7 | 0.2 ± 0.1 |
LETTER (26 actions) | 21.7 ± 0.7 | 14.3 ± 0.8 | 14.1 ± 0.5 | 12.4 ± 0.3 | 34.8 ± 5.4 | 37.0 ± 0.9 | 32.1 ± 1.0 | 2.8 ± 0.5 | |
Warfarin | LR (3 actions) | 6.9 ± 0.9 | 15.6 ± 18.7 | 9.0 ± 2.6 | 7.8 ± 0.8 | 24.3 ± 19.9 | 26.1 ± 6.8 | 7.5 ± 0.9 | 7.0 ± 0.7 |
LR (5 actions) | 10.0 ± 0.6 | 11.6 ± 6.7 | 9.3 ± 2.3 | 10.1 ± 0.9 | 12.7 ± 11.2 | 13.3 ± 5.4 | 9.3 ± 4.5 | 10.3 ± 0.6 | |
PHARMA (3 actions) | 20.6 ± 3.2 | 20.9 ± 18.7 | 17.5 ± 4.6 | 15.3 ± 1.1 | 19.0 ± 12.4 | 17.9 ± 4.1 | 15.1 ± 1.3 | 12.0 ± 0.7 | |
PHARMA (5 actions) | 11.5 ± 1.3 | 12.8 ± 5.0 | 13.6 ± 3.5 | 12.5 ± 1.1 | 9.6 ± 5.2 | 6.7 ± 4.2 | 8.0 ± 2.2 | 11.6 ± 0.6 | |