A new cross-validation technique to evaluate quality of recommender systems

Abstract

The topic of recommender systems is rapidly gaining interest in the user-behaviour modeling research domain. Over the years, various recommender algorithms based on different mathematical models have been introduced in the literature. Researchers interested in proposing a new recommender model or modifying an existing algorithm should take into account a variety of key performance indicators, such as execution time, recall and precision. Till date and to the best of our knowledge, no general cross-validation scheme to evaluate the performance of recommender algorithms has been developed. To fill this gap we propose an extension of conventional cross-validation. Besides splitting the initial data into training and test subsets, we also split the attribute description of the dataset into a hidden and visible part. We then discuss how such a splitting scheme can be applied in practice. Empirical validation is performed on traditional user-based and item-based recommender algorithms which were applied to the MovieLens dataset.