Classifying Limited Resource Data Using Semi-supervised SVM

Abstract

The success of machine learning models is subjected to the availability of sufficient amounts of training data. Avail- A bility of large amounts of labeled data is difficult especially in the domains of image, speech, video etc. Semi-supervised learning is an approach that uses additionally availableunlabeled data to improve performance of a model with limited data. To gain better performance with limited training data we suggest semi-supervised SVM models for EEG Signal classification and image classification tasks. We explore multiple approaches to semi-supervised learning based on Support vector machines: Semi-supervised Support vector machines (S3VM), S3VMlight, SVM-light and label switching based Support Vector Machine (Lap-SVM) for different tasks. Our experiments show that semi-supervised approaches when trained with sufficient unlabeled data can significantly improve performance of the model when compared with its counterpart supervised model. The proposed models are verified on 3 different benchmark data sets. Proposed semi-supervised approach for image classification task show a remarkable 20% improvement over baseline SVM model.