Persuasive-oriented Explanation Generation and Evaluation of Personalized Recommendation

Abstract

Persuasion is essential to explainable recommendation systems in making users accept the recommended items. However, current explainable recommendation models do not specifically strengthen the persuasion and lack an effective method to evaluate the persuasion. By defining the pre-review set, this work clearly defines the persuasion and designs an off-line evaluation method of persuasion without manual annotation, and constructs persuasion-enhanced data sets. Furthermore, this work develops a persuasion enhancement explainable recommendation model based on persuasion-enhanced data sets and cross-attention networks. This model not only improves the recommendation performance but also generates a high-quality natural language explanation. Extensive experiments verify the advantages of the proposed model in recommendation and explanation compared with the SOTA models.