Table 2.
DeepSurv and RSF hyperparameters search space value ranges used by the Optuna library to find the best combination that maximizes the test c-index value for both models.
| Hyperparameter | DeepSurv values | Hyperparameter | RSF values |
|---|---|---|---|
| Activation | Sigmoid, ReLU, SeLU | Number of trees | 100–1000; steps = 100 |
| Layers | 1, 2, 3, 4 | Max features | sqrt, log2 |
| Units | 8, 16, 32 | Max depth | 3–10; steps = 1 |
| Init method | Glorot, uniform | Sample size percentage | 0.60–0.85; step = 0.05 |
| Optimizer | Adam, SGD | Importance mode | Impurity corrected, permutation, |
| Learning rate | 10-5, 10-1 | Normalized permutation | |
| L2 reg | 10-6, 10-4 | ||
| Dropout | 0.1, 0.2, 0.4 |
For DeepSurv Batch Normalization is fixed to True, Batch and Dropout to False, and the Number of Epochs to 1500; for RSF the Min node size if fixed to 10, and the Number of Epochs to 1000.