Learning From Labeled And Unlabeled Data

Abstract

A common way to learn a classier is to train a classier with the help of available setof labeled information with predefine set of classes. Supervised learning algorithmsrequire set of labeled data of every class to engender a classification function. Oneof the shortcomings of this classical paradigm is that in order to learn the functionaccurately, a bulk of labeled examples are needed. To achieve accurate result pullof labeled data is required, what to do if only few labeled data is available? Task isto examine the influence on the accuracy of the recommender when it is built usingunlabeled examples in addition to the labeled examples.

Co-training and Self-Training allows to incorporate unlabeled examples while learning a classifier/recommender. Usefulness of these two algorithms is investigatedby experimental study using three different datasets those are MovieLens Dataset, Jester Dataset and Hetrec Dataset. Accuracy and f-measure are used as the evaluation measures.