An Ensemble Approach for Semi-Supervised Self Learning Algorithm

Machine learning methods are divided into two different groups, supervised learning and unsupervised learning, according to the presence of label information in the datasets. While unlabeled data is easy and cheap to access, labeling is difficult and time-consuming. Given the increasing amount of data today, labeling has become costly. Therefore, a semi-supervised learning approach is proposed. The self-learning algorithm, one of the oldest algorithms in this approach, starts learning/training with labeled samples. The class prediction probabilities of the trained model on unlabeled samples are obtained and pseudo labels are generated based on the predefined probability condition. This process continues iteratively until the unlabeled data is exhausted. The proposed method aims to save the model that trained at the end of each iteration and use it in subsequent iterations. In the experimental studies, the prediction result of the newly trained model and the average prediction results of the models saved from previous iterations are calculated. The ensemble result is obtained by averaging these two prediction results. Considering the preliminary findings obtained experimentally under certain constraints, the proposed method outperformed the semi-supervised self-learning algorithm by 3.42% on average in terms of accuracy metric.