Run iterative Semi-Supervised labeling experiment

[bdzyubak]

Use or simulate a dataset with only 30% of the data labeled. Run iterative semi-supervised labeling by the model to illustrate (expected) gain in performance vs training on only the original 30% of the labels.

[bdzyubak]

Due to the dataset being too easy (>0.96% accuracy), the amount of labeled data had to be reduced to 10%. Otherwise, differences between settings were in the noise and varied substantially between retraining runs.

Implemented experiment using SVM. A 0.9 probability threshold gives the highest accuracy, with two iterations needed to reach the accuracy plateau. These hyperparameters may not generalize to different datasets. Ideally, this would be run on a much more difficult dataset.
torch-control\projects\MachineLearning\semi_supervised_breast_cancer_classification\semi_supervised_svm.py

Semi-supervised learning with a DecisionTree/Random Forest/XGBoost would be interesting. It requires experimenting with tree depth as having arbitrary depth yields 100% prediction confidences. Pushing this to separate backlogged issue.