Abstract in Englisch:
Recently, the increasing complexity of Recommender Systems (RS) algorithms has led RS researchers to focus on improving user-centric aspects beyond algorithmic accuracy. In this context, explanations and opportunities for user control have been shown to increase overall system effectiveness and acceptance among RS users. In the RS literature, many attempts have been made to explain recommendations, mostly by exploiting ratings given by the user community for products or other objects (generally:items), in the context of Collaborative Filtering (CF), or by exploiting product features in Content-Based Filtering (CB). While these methods may suffice in simple recommendation scenarios, the complexity of other situations may require the use of several different data sources, for example depending on the product domain. To this end, hybrid systems combining the advantages of CF and CB techniques have been developed, but mainly to improve the algorithmic performance of RS. However, the relationship between recommended items and user preferences for product features has not been sufficiently exploited to generate explainable recommendations. Moreover, in most cases, item rating or re-rating is the only way for users to indicate their preferences and control the recommendation process. This thesis assumes that exploiting users’ feature preferences together with item ratings in a CF approach can generate more explainable recommendations and provide users with better opportunities to control the recommendation process which can be especially beneficial in complex domains such as digital cameras. From this starting point, this thesis first introduces the Feature-based Collaborative Filtering approach in the domain of digital cameras, which extends the conventional CF method by exploiting the feature preferences of similar users instead of item preferences with the goal of generating explainable recommendations. The first prototype implemented based on the approach was then empirically evaluated in two user studies. The results show that our novel approach is superior to conventional item-based CF in terms of subjective assessment of explanations, while a pure CB approach was perceived similarly positively. However, overall, participants appreciated the availability of extended explanations compared to conventional recommender approaches. Based on the findings of the user studies, we developed an extended version of the approach and implemented a prototype called Featuristic, in which we integrated interactive mechanisms into three main phases of the recommendation process: 1) Preference Elicitation, 2) Recommendations, and 3) Explanations – to improve explanation quality and user control. The Featuristic prototype was then evaluated in two user studies. The results showed that the implemented interactive mechanisms in several components of the recommendation process overall improved the explainability and controllability of RS compared to systems offering only non-interactive recommendations with limited or no explanations at all.