On Similarity Measures for a Graph-Based Recommender System


Kurt Z., BİLGE A., ÖZKAN K., Gerek O. N.

25th International Conference on Information and Software Technologies (ICIST), Vilniaus, Lithuania, 10 - 12 October 2019, vol.1078, pp.136-147 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 1078
  • Doi Number: 10.1007/978-3-030-30275-7_11
  • City: Vilniaus
  • Country: Lithuania
  • Page Numbers: pp.136-147
  • Keywords: Bipartite graph, Complex domain, Similarity measures, LINK-PREDICTION
  • Akdeniz University Affiliated: No

Abstract

Recommender systems are drawing increasing attention with several unresolved issues. These systems depend on personal user preferences on items via ratings and recommend items based on choices of similar users. A graph-based recommender system that has ratings of users on items can be shown as a bipartite graph in which vertices match users and items nodes, and edges correspond to ratings. Recommendation generation in a bipartite graph can be moderated as a sub-problem of link prediction. In the relevant literature, modified link prediction methods are employed to differentiate between fundamental relational dualities of like vs. dislike and similar vs. dissimilar. However, the similarity relationships between users/items are often ignored. We propose a new model that utilizes user-user and item-item similarity values with relational dualities in order to improve coverage and hits rate by carefully incorporating similarities. We compare five similarity measures in terms of hits rate and coverage while providing top-N recommendations. We scrutinize how such similarity measures perform with top-N item recommendation processes over the standard MovieLens Hetrec and MovieLens datasets. The experimental results show that hits rate and coverage can be improved by about 7% and 4%, respectively, with Jaccard and Adjusted-Cosine similarity measures being the best performing similarity measures. Significant differences/improvements are observed over the previous CORLP approach.