Review based emotion profiles for cross domain recommendation

摘要

Several e-commerce sites are reaping the benefits of Cross-Domain Recommendation (CDR) systems to cross-sell products, guide new users and increase revenues. Current research works augment user-item ratings with a variety of auxiliary information such as location, personality, geo-tags and multimedia content that link multiple domains to provide effective CDR. In this paper, we propose a fresh perspective for generating recommendations across different domains by tapping the emotions that are encapsulated within user generated textual content such as reviews, blogs and comments. Such emotions serve as strong socio-psychological links between various entertainments domains and have the potential to obviate the cold start problems. Our CDR scheme uses an enriched emotion lexicon to analyze the emotions in online content expressed by users in the source and target domains and generates emotion-profiles of items and users in both domains. Subsequently, it applies collaborative filtering to match these profiles in order to recommend items in the target domain. We illustrate the working of our emotion-based CDR scheme using the movie and book domains as a case study. Experimental results on Movielens and Bookcrossing datasets yield 28.9% F1-measure which is a marked improvement of 71.1% as compared with a recently reported topic modeling approach to CDR for entertainment domains.