An Empirical Study on Effect of Semantic Measures in Cross-Domain Recommender System in User Cold-Start Scenario

Abstract

Cold-start problem is one of the fatal problems in recommender system. The development of cross-domain recommender system (CDRS) provides feasibility to deal with the problem, while it needs to handle heterogeneous information when linking different domains. Some existing semantic measures based on Linked Open Data (LOD) are likely to play a positive role in this research area. According to the different LOD information used, we analyze two classes of semantic measures (similarity and relatedness) to explore the relationship between LOD-based domain correlation and user interests, and study the performance of different semantic measures on a cross-domain recommendation framework. Through experiments on a real-world dataset, this work has identified that the similarity measures can accurately capture the user’s existing interests while the relatedness measures can produce diverse recommendations. Besides, by comparing with some representative methods, the experiment demonstrated that the cross-domain recommendation method could provide users with satisfactory recommendations even in the cold-start scenario.