Table 1 Default models and the possible parameter choices for each model under different conditions
Classifier | Non-tunable parameters | Number of training features (≥ 4 features) | Parameters (If number of samples > 6) | Parameters (if number of samples ≤ 6) |
|---|---|---|---|---|
Logistic regression (LBFGS solver) | max_iter = 2000 penalty = “l2” | Randomly selected—user defined | A grid search using fivefold stratified cross-validation is used to choose C from logarithmically spaced values in the range of 10–4 and 104 | C parameter set to 1.0 |
Logistic regression (Liblinear solver) | max_iter = 2000 penalty = “l1” | Full | A grid search using fivefold stratified cross-validation is used to choose C from logarithmically spaced values in the range of 10–4 and 104 | C parameter set to 1.0 |
Linear SVC | max_iter = 2000 | Randomly selected—user defined | A grid search using fivefold stratified cross-validation is used to choose alpha (for SGD classifiers) or C (for the linear SVC). The possible choices for these parameters are 0.001, 0.01, 0.1, 1.0, 10, 100. In the case of the SGD Classifier, the loss function (hinge or modified Huber) is also chosen using 5 cross-validation | C parameter set to 1.0 |
Stochastic gradient descent classifier (L2 penalty) | max_iter = 2000 | Randomly selected—user defined | A grid search using fivefold stratified cross-validation is used to choose alpha (for SGD classifiers) or C (for the linear SVC). The possible choices for these parameters are 0.001, 0.01, 0.1, 1.0, 10, 100. In the case of the SGD Classifier, the loss function (hinge or modified Huber) is also chosen using 5 cross-validation | Alpha parameter set to 1.0, loss function (hinge or modified Huber) is randomly chosen |
Stochastic gradient descent classifier (elastic-net penalty) | max_iter = 2000 | Full | A grid search using fivefold stratified cross-validation is used to choose alpha (for SGD classifiers) or C (for the linear SVC). The possible choices for these parameters are 0.001, 0.01, 0.1, 1.0, 10, 100. In the case of the SGD Classifier, the loss function (hinge or modified Huber) is also chosen using 5 cross-validation | Alpha parameter set to 1.0, loss function (hinge or modified Huber) is randomly chosen |
Ridge regression | NA | Randomly selected—user defined | Alpha chosen from logarithmically spaced values in the range of 10–3 and 104 using generalized cross validation | NA |
Neural network | batch_size = 32 epochs = 300 validation_split = 0.10 min_delta = 0.0001 patience = 40 See text for architecture details | Randomly selected—user defined | NA | NA |