An ontology-based recommender system using scholar's background knowledge

Abstract

Scholar’s recommender systems recommend scientific articles based on the similarity of articles to scholars’ profiles, which are a collection of keywords that scholars are interested in. Recent profiling approaches extract keywords from the scholars’ information such as publications, searching keywords, and homepages, and train a reference ontology, which is often a general-purpose ontology, in order to profile the scholars’ interests. However, such approaches do not consider the scholars’ knowledge because the recommender system only recommends articles which are syntactically similar to articles that scholars have already visited, while scholars are interested in articles which contain comparatively new knowledge. In addition, the systems do not support multi-area property of scholars’ knowledge as researchers usually do research in multiple topics simultaneously and are expected to receive focused-topic articles in each recommendation. To address these problems, this study develops a domain-specific reference ontology by merging six Web taxonomies and exploits Wikipedia as a conflict resolver of ontologies. Then, the knowledge items from the scholars’ information are extracted, transformed by DBpedia, and clustered into relevant topics in order to model the multi-area property of scholars’ knowledge. Finally, the clustered knowledge items are mapped to the reference ontology by using DBpedia to create clustered profiles. In addition a semantic similarity algorithm is adapted to the clustered profiles, which enables recommendation of focused-topic articles that contain new knowledge. To evaluate performance of the proposed approach, three different data sets from scholars’ information in Computer Science domain are created, and the precisions in different cases are measured. The proposed method, in comparison with the baseline methods, improves the average precision by 6% when the new reference ontology along with the full scholars’ knowledge is utilized, by an extra 7.2% when scholars’ knowledge is transformed by DBpedia, and further 8.9% when clustered profile is applied. Experimental results certify that using knowledge items instead of keywords for profiling as well as transforming the knowledge items by DBpedia can significantly improve the recommendation performance. Besides, the domain-specific reference ontology can effectively capture the full scholars’ knowledge which results to more accurate profiling.