Conformal recommender system

Abstract

Conformal prediction is a relatively recent approach for quantifying the uncertainty in classification problems. It can provide reliable measures of confidence for predicting class labels of unclassified patterns. This framework is applicable to classification problems but the implementation of conformal prediction for classification depends on the classification algorithm at hand. In literature, several classification algorithms are used to incorporate the framework of conformal prediction. In this paper, we extend the concept of conformal prediction to recommender systems and propose a Conformal Recommender System (CRS). We define nonconformity measure, a key concept of conformal prediction, for recommender system and show that it satisfies the exchangeability property. We also show that our proposed conformal recommender system satisfies the desirable properties of conformal prediction such as validity and efficiency. With this we are in a position to build a better recommender system. We compare our method with 12 state-of-the-art recommender algorithms on 10 different datasets to corroborate this claim.