Generating Personalized Explanations for Recommender Systems Using a Knowledge Base

Introduction

Over the years, there has been an increase in recommender systems being implemented in various applications. Additionally, an increase in information has partially occurred due to the rise of the internet. Obtaining desired information from vast data sources is difficult due to the scale and rapid growth of data. A recommender system could be considered one way to address information overload and leverage this information to generate efficient recommendations. However, a significant concern for the recommendation systems has been a lack of explainability. We already know that providing explanations of recommendations increases user trust and improves effectiveness, efficiency, and satisfaction (Tintarev and Masthoff 2007; Zhang and Chen 2020; Kouki et al. 2019; Cleger-Tamayo et al. 2012; Fu et al. 2020). While striving for interpretable recommendations, a trade-off between accuracy and interpretability exists. Since we focus on adapting a model-agnostic approach, we generate explanations for the recommendations regardless of the method implemented to generate recommendations. This approach offers more flexibility to generate explanations while relying on additional features other than the features implemented to generate recommendations.

Transparency and justification are the two primary purposes of explanations, and they can be implemented in either a model-agnostic or model-intrinsic environment. Transparency enables users to understand the system and improves the clarification of the recommendation process. Figure 1 shows how a model-intrinsic approach integrates the explanations process into a recommendation model. However, since the model-intrinsic approaches prioritize accuracy over explainability, the generated explanations may not be as reliable. On the other hand, the model-agnostic approaches in Figure 2 provide justifications for the generated explanations. The recommendation module in the model-agnostic approaches is considered a black box, and the explanations can be generated using an independent method. However, the explanations lack persuasiveness since it only relies on recommendations obtained from the Recommender System in Figure 2.

The model-agnostic method is crucial since it generates explanations for the recommended items regardless of the implemented approach to generate recommendations. In this work, we extend our previous work (Chen and Miyazaki, 2020) to provide personalized explanations for recommended items and conduct a user study for evaluating our personalized explanations. In this study, we propose a Personalized Explanation Generation System, PerExGen (PersonalizedExplanationGeneration), which generates personalized explanations for recommender systems using a model-agnostic approach. We assume that the dataset information is not sufficient, and there is a need to generate high-quality personalized explanations. We make use of a knowledge graph to generate high-quality personalized explanations. Knowledge graphs make it easier to integrate with other approaches to generate explanations.

We use a Task-Specialized Item Knowledge Graph to exploit the available information to find relationships between the users and the items and extract paths that can be used to generate explanations. Our proposed system consists of the recommendation and explanation modules. We intend to focus more on the explanation mechanism and construct a Task-Specialized Item Knowledge Graph to generate personalized explanations. We extract relevant information, rank the extracted candidate paths, and generate natural language personalized explanations using templates. Our system aims to generate personalized explanations but can also generates general explanations for general users.

The contributions of our work are summarized here: we (1) propose a personalized explanation generation system PerExGen that uses a publicly available knowledge base and (2) introduce a Task-Specialized Item Knowledge Graph and generate personalized explanations for a recommendation by employing a model-agnostic approach, and (3) evaluate our system by conducting a user study and show that our proposed approach generated more persuasive and natural explanations than the explanations from the existing systems.

Figure 1.

A Model-agnostic Approach