PROJECT TITLE :

Semisupervised Multiple Choice Learning for Ensemble Classification

ABSTRACT:

Due to the fact that it is so effective at enhancing the predictive performance of classification models, Ensemble Learning has a wide variety of applications that have been met with great success. In this article, we propose an approach known as semisupervised multiple choice learning (SemiMCL) for jointly training a network ensemble using data that is only partially labeled. Our model places a primary emphasis on enhancing a labeled data assignment among the constituent networks and utilizing unlabeled data to better capture domain-specific information. This is done in order to effectively facilitate semisupervised classification, which is the model's ultimate goal. In contrast to more traditional learning models based on multiple choice questions, the constituent networks acquire knowledge of multiple tasks during the training process. In particular, a secondary reconstruction task is incorporated in order to acquire knowledge of domain-specific representation. When minimizing the conditional entropy with respect to the posterior probability distribution, we use a negative l1 -norm regularization in order to achieve our goal of performing implicit labeling on reliable unlabeled samples. This allows us to achieve the best possible results. In order to prove that the proposed SemiMCL model is both effective and superior to other models, a large number of experiments are run on a variety of datasets that come from the real world.