Review-based Explanations for Recommendations in E-Commerce

• Applicants: Professor Dr. Mathias Klier; Professor Dr. Steffen Zimmermann
• Subject Area: Operations Management and Computer Science for Business Administration
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 544814728

Recommender systems can mitigate information overload on e-commerce platforms by recommending products or services aligned with users' preferences. State-of-the-art recommender systems use artificial intelligence (AI) methods and incorporate customer reviews to achieve high recommendation quality. However, these systems often represent 'black boxes' to users. As a result, trust in these systems and the willingness to follow their recommendations is limited among many users. The emerging field of Explainable Artificial Intelligence (XAI) aims to automatically generate explanations for decisions and functioning of AI systems. User-centric explanations can increase trust by assisting users of recommender systems in critically evaluating recommendations. Furthermore, the integration of reviews into explanations holds the potential to further increase user trust, as reviews are considered a trustworthy information source. Yet, there are no existing methods capable of generating user-centric explanations for recommender systems while incorporating reviews. Therefore, the planned project 'Review-based Explanations for Recommendations in E-Commerce' addresses the following research question: How can review-based explanations be automatically generated that disclose the formation of recommendations from review-based recommender systems in order to increase trust and user acceptance? This research project falls within the field of design-oriented Information Systems Research and comprises two sub-projects, T1 and T2. In T1, a new user-centric explanation method is developed, which generates review-based explanations for recommendations from recommender systems. This involves developing three core components: structuring of reviews around concepts relevant to humans (core component 1), systematic variation of reviews based on these relevant concepts (core component 2), and determination of influences on recommendations through this systematic variation of reviews and generation of user-centric explanations (core component 3). In T2, the explanation method is evaluated both functionally-grounded and human-grounded with regard to multiple criteria including user trust and user acceptance and further refined through design cycles. Finally, the explanation method is evaluated application-grounded in a field experiment in order to examine the added value of the method with regard to real (economic) user behavior (e.g., acceptance of the recommendations). The outcomes of these intertwined sub-projects include a novel explanation method generating review-based explanations for recommendations of recommender systems, along with insights into its value and impacts on user behavior.

• DFG Programme: Research Grants