Diana Carolina Hernandez Bocanegra: Argumentative Explanations for Recommendations Based on Reviews (Dr.-Ing.)

Description

Recommender systems (RS) assist users in making decisions on a wide range of tasks, while preventing them from being overwhelmed by enormous amounts of choices. RS prevalence is such that many users of information-based technologies interact with them on a daily basis. However, many of these systems are still perceived as black boxes by users, who often have no way of seeing or requesting the reasons why certain items are recommended, potentially leading to negative attitudes towards RS by users. Providing explanations in RS can bring several advantages for users’ decision making and overall user experience. Although different explanatory approaches have been proposed so far, the general lack of user evaluation, and validation of concepts and implementations of explainable methods in RS, have left open many questions, related to how such explanations should be structured and presented. Also, while explanations in RS have so far been presented mostly in a static and non-interactive manner, limited work in explainable artificial intelligence have emerged addressing interactive explanations, enabling users to examine in detail system decisions. However, little is known about how interactive interfaces in RS should be conceptualized and designed, so that explanatory aims such as transparency and trust are met. This dissertation investigates interactive, conversational explanations that enable users to freely explore explanatory content at will. Our work is grounded on RS explainable methods that exploit user reviews, and inspired by dialog models and formal argument structures. Following a user-centered approach, this dissertation proposes an interface design for explanations as interactive argumentation, which was empirically validated through different user studies. To this end, we implemented a RS able to provide explanations both through a graphical user interface (GUI) navigation and a natural language interface. The latter consists of a conversational agent for explainable RS, which supports conversation flows for different types of questions written by users in their own words. To this end, we formulated a model to facilitate the detection of the intent expressed by a user on a question, and collected and annotated a dataset helpful for intent detection, which can facilitate the development of explanatory dialog systems in RS. The results reported in this dissertation indicate that providing interactive explanations through a conversation, i.e. an exchange of questions and answers between the user and the system, using both GUI-navigation or natural language conversation, can positively impact users evaluation of explanation quality and of the system, in terms of explanatory aims like transparency, and trust.

Published as