SemiScope study: Tuning the classifier alone recovers 86% of SSL gains

A new paper on arXiv (2607.00113) tackles a common pitfall in applying semi-supervised learning (SSL) to security classification: researchers often treat SSL pipelines as black boxes, using default parameters and fixed classifiers. The authors, led by Rui Shu, introduce SemiScope—an analysis instrument that disentangles whether the reported gains from joint optimization of SSL pipelines come from actually improving SSL-classifier interactions or simply from better tuning the downstream classifier.

Using five binary tabular security datasets and classical SSL with tree-based classifiers, SemiScope employs Bayesian optimization to jointly tune SSL settings, confidence filtering, oversampling, and the classifier. The key control, called Tuned-Clf, fixes SSL parameters to defaults but allocates the same 100-trial classifier budget and validation-set threshold tuning as the full pipeline. At just 10% labeled data, SemiScope outperforms all default SSL baselines by 0.7 to 12.7 g-measure points. However, under equal budget, Tuned-Clf is statistically equivalent to the full pipeline on four of five datasets. Classifier hyperparameter optimization alone recovers a median 86% of SemiScope's gain over default self-training plus random forest. The authors conclude with a simpler, deployment-ready recipe: use self-training, tune the classifier with Bayesian optimization, and adjust the decision threshold on validation data. This approach reaches within 1 g-measure of a fully supervised random forest using only 20–30% labels on four datasets (40% on Drebin).