Certain investigation on constrained Semi supervised learning

Abstract

Semi-Supervised Learning (SSL) is a learning framework which newlineleverages unlabeled data, when labels are hard to obtain. These methods are newlineused to learn the structure of data from the unlabeled instances, without newlineconsidering the need of labels. It is preferred when compared to supervised newlineand unsupervised methods as it reduces the human effort and increases the newlineaccuracy. Since it lies between supervised and unsupervised learning, it newlinecombines the tasks of both the above methods. SSL is getting more newlinesignificance in recent years due to the availability of large volumes of data newlineand limitations of getting labels for the data. These are relevant to scenarios newlinewhere there is a scarcity of labels particularly in application domains of newlinemedicine, discovery of drugs, automated diagnosis etc. Due to the constraints newlinein getting labeled data, SSL methods are widely used in applications. newlineExtensive usage of SSL finds their limitations in practice. The proposed newlineresearch work attempts to find better alternatives for the limitations identified. newlineThis research focuses on two constraints of SSL. newlineFeature Selection for Semi-Supervised Learning newlineSemi-Supervised Learning for unseen classes. newlineThe task of newlineknown as feature selection, which is one of the widely used techniques for newlinepattern analysis and data mining. It is a significant task that selects relevant newlinefeatures and removes irrelevant features to improve the performance of newlinelearner. Semi-supervised feature selection makes use of both labeled and newlineunlabeled data simultaneously to evaluate the feature relevance. newline