With the popularity of Linked Open Data (LOD) and the associated rise infreely accessible knowledge that can be accessed via LOD, exploiting LOD forrecommender systems has been widely studied based on various approaches such asgraph-based or using different machine learning models with LOD-enabledfeatures. Many of the previous approaches require construction of an additionalgraph to run graph-based algorithms or to extract path-based features bycombining user- item interactions (e.g., likes, dislikes) and backgroundknowledge from LOD. In this paper, we investigate Factorization Machines (FMs)based on particularly lightweight LOD-enabled features which can be directlyobtained via a public SPARQL Endpoint without any additional effort toconstruct a graph. Firstly, we aim to study whether using FM with theselightweight LOD-enabled features can provide competitive performance comparedto a learning-to-rank approach leveraging LOD as well as other well-establishedapproaches such as kNN-item and BPRMF. Secondly, we are interested in findingout to what extent each set of LOD-enabled features contributes to therecommendation performance. Experimental evaluation on a standard dataset showsthat our proposed approach using FM with lightweight LOD-enabled featuresprovides the best performance compared to other approaches in terms of fiveevaluation metrics. In addition, the study of the recommendation performancebased on different sets of LOD-enabled features indicate that property-objectlists and PageRank scores of items are useful for improving the performance,and can provide the best performance through using them together for FM. Weobserve that subject-property lists of items does not contribute to therecommendation performance but rather decreases the performance.
展开▼
