Linked open data-based explanations for transparent recommender systems. Issue 121 (January 2019)

• Record Type: Journal Article
• Authors: Musto, Cataldo; Narducci, Fedelucio; Lops, Pasquale; de Gemmis, Marco; Semeraro, Giovanni
• Is Part Of: International journal of human-computer studies. Issue 121(2019)
• Page Start: 93
• Page End: 107
• Publication Date: 2019-01
• Subjects: Linked open data -- Explanation -- Recommender systems -- User interface -- User study; Human-machine systems -- Periodicals; Systems engineering -- Periodicals; Human engineering -- Periodicals; Human engineering; Human-machine systems; Systems engineering; Periodicals; Electronic journals; 004.019
• DOI: 10.1016/j.ijhcs.2018.03.003
• Languages: English
• ISSNs: 1071-5819

Abstract

Highlights: We design the main components of an algorithm-agnostic framework to generate natural language explanations ; We propose a methodology to extract descriptive (direct and indirect) properties about the items, and we use these properties to feed a graph-based explanation model; We define a scoring function to rank these explanation patterns and we use the most relevant ones to generate a template-based natural language explanation; We validate our methodology by carrying out a large user study (N = 680) in three different domains, as movies, books and music; We integrate our methodology in a conversational recommender system implemented as a Telegram Bot. Abstract: In this article we propose a framework that generates natural language explanations supporting the suggestions generated by a recommendation algorithm. The cornerstone of our approach is the usage of Linked Open Data (LOD) for explanation aims. Indeed, the descriptive properties freely available in the LOD cloud (e.g., the author of a book or the director of a movie) can be used to build a graph that connects the recommendations the user received to the items she previously liked via the properties extracted from the LOD cloud. In a nutshell, our approach is based on the insight that properties describing the items the user previously liked as well as the suggestions she received can be effectively used to explain the recommendations. Such a framework is both algorithm-independent and domain-independent, thus it can generate a natural language explanation for every kind of recommendation algorithm, and it can be used to explain a single recommendation ( Top-1 scenario) as well as a group of recommendations ( Top-N scenario). It is worth noting that the algorithm-independent characteristic does not mean that the framework is able to explain to the user how the recommendations have been generated and how the recommendation algorithm works. The framework explains to users why they might like the recommended items, independently from the recommendation algorithm that generated the recommendations. In the experimental evaluation, we carried out a user study (N = 680) aiming to investigate the effectiveness of our framework in three different domains, as movies, books and music . Results showed that our technique leads to transparent explanations for all the domains, and such explanations resulted independent of the specific recommendation algorithm in most of the experimental settings. Moreover, we also showed the goodness of our strategy when an entire group of recommendations has to be explained. As a case study, we integrated the framework in a real-world application, a conversational recommender system implemented as a Telegram Bot. The idea is to use the explanation for supporting both the training phase (when the user expresses her preferences) and the recommendation step (when the user receives the recommendations). Interesting outcomes emerge from these preliminary experiments.