Table 1 RMSE and UQ evaluation metrics for the 9 RF models trained on Crippen’s logP from [15]. The simulated values, \(NLL^{sim}\) and \(\rho _{rank}^{sim}\) is the average of 1000 simulated sets of test errors based on the predicted uncertainties. The number in parenthesis is the standard deviation of the 1000 values
\(N_{train}\) | RMSE | \(R^2\) | a | b | \(\rho _{rank}\) | \(\rho _{rank}^{sim}\) | \(A_{mis}\) | NLL | NLL\(^{sim}\) |
|---|---|---|---|---|---|---|---|---|---|
100 | 1.29 | 0.84 | 0.62 | 0.62 | 0.11 | 0.19 (0.01) | 0.05 | 1.73 | 1.46 (0.01) |
500 | 1.09 | 0.85 | 0.64 | 0.45 | 0.11 | 0.19 (0.01) | 0.03 | 1.51 | 1.39 (0.01) |
1000 | 1.01 | 0.85 | 0.55 | 0.45 | 0.10 | 0.19 (0.01) | 0.00 | 1.42 | 1.40 (0.01) |
5000 | 0.93 | 0.81 | 0.57 | 0.42 | 0.10 | 0.18 (0.01) | 0.01 | 1.35 | 1.29 (0.01) |
10,000 | 0.90 | 0.82 | 0.58 | 0.40 | 0.11 | 0.19 (0.01) | 0.01 | 1.32 | 1.24 (0.01) |
20,000 | 0.86 | 0.86 | 0.58 | 0.37 | 0.11 | 0.18 (0.01) | 0.01 | 1.26 | 1.21 (0.01) |
50,000 | 0.81 | 0.88 | 0.61 | 0.31 | 0.11 | 0.19 (0.01) | 0.00 | 1.19 | 1.16 (0.01) |
100,000 | 0.77 | 0.85 | 0.67 | 0.26 | 0.13 | 0.20 (0.01) | 0.00 | 1.15 | 1.12 (0.01) |
150,000 | 0.75 | 0.91 | 0.69 | 0.23 | 0.14 | 0.21 (0.01) | 0.01 | 1.11 | 1.09 (0.01) |