A Bisociated Domain-Based Serendipitous Novelty-Recommendation Technique for Recommender Systems

Abstract

Traditional recommendation paradigms such as content-based filtering (CBF) tend to recommend items that are very similar to user profile characteristics and item input, resulting in the classical twin problem of overspecialization and concentration bias of recommendations. This twin problem is prevalent with CBF recommender systems due to the utilisation of accuracy metrics to retrieve similar items, and, limiting recommendation computations to single recognized user-centered domains, rather than cross-domains.  This paper proposes a Bisociated domain-based serendipitous novelty recommendation techniques using Bisolinkers exploratory creativity discovery technique. The use of Bisolinkers enables establishing unique links between two seemingly unrelated domains, to enhance recommendation accuracy and user satisfaction. The presence of similar terms in two habitually incompatible domains demonstrates that two seemingly unrelated domains contain elements that are related and may act as a link to connect these two domains.

Keywords: recommender systems, novelty, machine learning, outlier detection, bisociation