Cross Domain Recommender Systems

Abstract

Cross domain recommender systems (CDRS) can assist recommendations in a target domain based on knowledge learned from a source domain. CDRS consists of three building blocks: domain, user-item overlap scenarios, and recommendation tasks. The objective of this research is to identify the most widely used CDRS building-block definitions, identify common features between them, classify current research in the frame of identified definitions, group together research with respect to algorithm types, present existing problems, and recommend future directions for CDRS research. To achieve this objective, we have conducted a systematic literature review of 94 shortlisted studies. We classified the selected studies using the tag-based approach and designed classification grids. Using classification grids, it was found that the category-domain contributed a maximum of 62%, whereas the time domain contributed at least 3%. User-item overlaps were found to have equal contribution. Single target domain recommendation task was found at a maximum of 78%, whereas cross-domain recommendation task had a minor influence at only 10%. MovieLens contributed the most at 22%, whereas Yahoo-music provided 1% between 29 datasets. Factorization-based algorithms contributed a total of 37%, whereas semantics-based algorithms contributed 6% among seven types of identified algorithm groups. Finally, future directions were grouped into five categories.