Accuracy Improvement of Neural Networks Under Semi-Supervised Learning
Victor M. Sineglazov, Olena Chumachenko, Kyrylo Lesohorskyi · 2023
This work concerns the problem of improving the accuracy of semi-supervised machine learning in solving a multiclass classification problem by choosing the optimal semi-supervised machine learning algorithm from a set of known ones. As the optimality criteria, we use the metrics proposed by the authors based on the criteria for the purity of clusters and normalized general information, which enables the assessment of cluster assumption for a given dataset. As an example of using the proposed approach, we consider the presence of the cluster assumption in the training sample, and, accordingly, the choice of the optimal semi-supervised learning algorithm in the presence of this assumption.