Data sets for offline evaluation of scholar's recommender system

Abstract

In an offline evaluation of recommender systems, data sets have been extensively used to measure the performance of recommender systems through statistical analysis. However, many data sets are domain and application dependent and cannot be engaged in different domains. This paper presents the construction of data sets for the offline evaluation of a scholar’s recommender system that suggests papers to scholars based on their background knowledge. We design a cross-validation approach to reduce the risk of false interpretations by relying on multiple independent sources of information. Our approach addresses four important issues including the privacy and diversity of knowledge resources, the quality of knowledge, and the timely knowledge. The resulting data sets represent the instance of scholar’s background knowledge in clusters of learning themes, which can be used to measure the performance of the scholar’s recommender system.