Table 6 The search space of each classifier based on the distributions over its hyperparameters (n.b. F denotes feature count; for biased categorical distributions, tuples (ps, v) designate the sampling probability and the value assigned)
From: A principled machine learning framework improves accuracy of stage II colorectal cancer prognosis
Classifier | Hyperparameter | Distribution | Values |
|---|---|---|---|
SVM, linear kernel | C | Log-uniform | [ln (1e−5), ln (1e2)] |
Class weight | Categorical | Balanced or none | |
SVM, RBF kernel | C | Log-uniform | [ln (1e−5), ln (1e2)] |
Gamma | Log-uniform | [ln (1e−3), ln (1e3)] | |
Class weight | Categorical | Balanced or none | |
LR | Type of penalty | Categorical | L1 or L2 |
C | Log-uniform | [ln (1e−5), ln (1e2)] | |
Class weight | Categorical | Balanced or none | |
RF | Number of trees | Log-uniform integer | [10, 1000] |
Criterion | Categorical | Gini or entropy | |
Maximum features | Biased categorical | (0.2, √F), (0.1, ln F), (0.1, F), (0.6,U(0, F)) | |
Maximum depth | Biased categorical | (0.1, 2), (0.1, 3), (0.1, 4), (0.7, none) | |
Bootstrap | Categorical | True or False | |
Class weight | Categorical | Balanced or none | |
KNN | K | Log-uniform integer | [1, 50] |
Weights | Categorical | Uniform, or Euclidean distance | |
Metric | Categorical | Balanced or none | |
P | Categorical | Balanced or none |
