Evaluation of Usefulness of Unlabeled Data in Learning a Recommender

Abstract

Supervised learning algorithms require labeled training examples from every class to engender a classification function. One of the shortcomings of this classical paradigm is that in order to learn the function accurately, a large number of labeled examples are needed. There are many situations (e.g. a new user in an online recommender system) where for every class,only a small set of labeled examples is available. Situations such as these encourage to investigate about the usefulness of unlabeled examples in learning a recommender. The main objective of this paper is to examine the influence on the accuracy of the recommender when it is built using unlabeled examples in addition to the labeled examples. Co-Training algorithm which allows to incorporate unlabeled examples while learning a classifier/recommender. Usefulness of this algorithm is investigated by means of experimental study using hetrec2011-movielens-2k data set.Accuracy and f-measure are used as the evaluation measures.