An Optimized Semi-Supervised Learning Approach for High Dimensional Datasets

The semi-supervised learning is one of the most interesting fields for research developments in the machine learning domain beyond the scope of supervised learning from data. Medical diagnostic process works mostly in supervised mode, but in reality, we are in the presence of a large amount of unlabeled samples and a small set of labeled examples characterized by thousands of features. This problem is known under the term “the curse of dimensionality”. In this study, we propose, as solution, a new approach in semi-supervised learning that we would call Optim Co-forest. The Optim Co-forest algorithm combines the re-sampling data approach (Bagging Breiman, 1996) with two selection strategies. The first one involves selecting random subset of parameters to construct the ensemble of classifiers following the principle of Co-forest (Li & Zhou, 2007). The second strategy is an extension of the importance measure of Random Forest (RF; Breiman, 2001). Experiments on high dimensional datasets confirm the power of the adopted selection strategies in the scalability of our method.