An integrated approach for transparent, multi-level decision support in interactive recommender systems

To facilitate the often complex product search on the Internet, various techniques such as search and filtering mechanisms or recommendation components are employed. Widely used faceted filtering systems allow users to successively narrow down the result set according to the filter settings that are applied. However, this method requires clear objectives and domain knowledge on the part of the users to purposefully refine the search results. Alternatively, conversational recommendation systems in the form of product advisors have recently gained significance. These systems recommend suitable products based on a series of questions, focusing more on the users' tasks and usage scenarios rather than the technical properties of the products. Thus, even users with little domain knowledge can find adequate products. Online environments frequently offer their users different decision aids, and it has been observed that users switch between them during their search. If the chosen decision aid seems ineffective, for example, because the user is overwhelmed with the available configuration options or, conversely, the chosen decision aid offers too little control, users switch to a decision aid that appears to be more helpful. However, this search behavior is typically not supported by environments because the provided tools are mostly isolated with no data exchange between them. As a result, all previous settings are lost and the search process has to start over. Consequently, users must re-enter previously set preferences, increasing cognitive load and making the search and decision-making process inefficient. To address this issue, this work proposes an integrated approach for an interactive recommender system that combines different decision-making components and allows users to freely switch between them. Selection actions in one component affect the other component in a consistent and transparent manner and can be further adjusted there, enabling an efficient search. We propose methods to enable a coupling of various components and allow users a seamless transition between them to foster multi-level decision support. Furthermore, we describe approaches for visualizing and explaining the relationships between the components, thereby enabling a transparent and comprehensible recommendation process. Findings from user studies examining user behavior and subjective evaluation through prototype implementations demonstrate that the proposed hybrid approach is suitable for supporting users in their search and decision-making process. It was shown that the combination of a conversational advisor with filtering mechanisms using a knowledge graph and the resulting explanation of the mutual influences, compared to traditional systems in which components work in isolation, positively impacts the perceived transparency of the recommendation process and the acceptance of the results.