Model-Agnostic Knowledge Graph Embedding Explanations for Recommender Systems

Explanations in recommender systems play an essential role in enhancing transparency, trust, and persuasiveness. In that regard, Knowledge Graphs (KGs) model-agnostic explanations do not rely on user-inputted data such as reviews or require any changes in a recommendation algorithm to provide explanations. The state-of-the-art of model-agnostic KG explainable algorithms are based on syntactic approaches that consider the trade-off of attributes among the user-interacted items and the catalog to explain recommendations. In this study, we propose a novel model-agnostic KG algorithm for explanations. Our approach utilizes KG embeddings to rank explanations based on the path’s similarity to the user. Specifically, we train an embedding algorithm on a KG and compare path embeddings, composed of node and edge embeddings, to the user embedding derived from previously interacted item embeddings. Our proposed method is evaluated by comparing it against three baselines representing the state-of-the-art of KG explanation algorithms. We assess explanation quality using three metrics: diversity and popularity of attributes displayed in explanations and recency of interacted items. Results indicate that the embedding approach achieves a superior balance between attribute popularity and explanation diversity. Furthermore, our analysis emphasizes the importance of tailored metrics for evaluating explanations in recommender systems.