An exploration of recommender systems explanation paradigms: generating and evaluating syntactic, semantic, and generative models with knowledge graphs

• Ano de defesa: 2024
• Autor(a) principal: Zanon, Andre Levi
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Explainable AI; Grafos de conhecimento; Inteligência artificial explicável; Knowledge graphs; Recommender systems; Sistemas de recomendação
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-20032025-103852/

Explainable recommender systems provide users with reasons to interact with items, enhancing transparency, persuasiveness, trust, efficiency, and satisfaction, thereby improving the overall user experience with the system. To generate such explanations with Knowledge Graphs (KG), algorithms choose between a set of explanation paths, which connect users interacted item nodes with a recommended item node, based on shared attributes. Three main paradigms are used to generate these paths: syntactic, semantic, and generative. Syntactic approaches select paths based on the number of links between item nodes; semantic approaches embed KG structures into a latent space; and generative models leverage language models to understand and generate explanations. However, the evaluation of explanations is often overlooked, with most studies relying on online experiments and remaining unclear how offline metrics align with user perception and explanation goals. This gap creates challenges in assessing state-of-the-art explanation algorithms and, consequently, in advancing the field. In this thesis, we design and evaluate explanation algorithms across syntactic, semantic, and generative paradigms, using both online and offline metrics to assess the different explanation paradigms and find correlations between user perception and explanation path selection. Our findings highlight that syntactic, semantic, and generative models are, respectively, an evolution in offline explanation metrics and that the diversity and popularity of attributes in explanation paths impact user perception. Finally, we provide guidelines for robust offline evaluation of explanations in recommender systems.