Abstract
Recommender systems are an application of artificial intelligence techniques where typically past behaviour of users is used to make predictions about their interests and to support them in identifying items they presumably like.
Over the past two decades and due to Internet's ever-growing plethora of choices recommender systems have received a lot of interest, both from industry and research, and yielded large progress in the field. However, most of the research is about predicting the user's preferences (i.e. precision-oriented), while focusing less on users' inherent needs and their perceptions of user-interfaces.
In this thesis we therefore digged deeper into users' perception of ratings, how they can be exploited to make recommendations explainable, how different characteristics of rating summarizations impact users' online choices and how they are additionally moderated by interpersonal differences.
We first researched how recommendation algorithms can exploit inherent information in interactions/ratings to provide explanations and/or to counteract the filter bubble.
The first contribution extends well-known recommendation algorithms based on matrix factorisation to trade-off performance with novelty and explainability of proposed items. Moreover, we defined a novel evaluation metric to measure the performance of the proposed method. Our extensive experimental results demonstrate that we attain high accuracy although recommending also more novel and explainable items.
Second, by running series of three conjoint experiments, we explore how these collaborative explanations that typically constitute summarizations of rating distributions (i.e., in the form of number of ratings, mean, variance, skewness, bimodality, or origin of the ratings) impact users' decisions. In the first study, with around 80 participants, we identified that users are primarily guided by the mean and the number of ratings. In the second study with over 200 participants, we confirm that users are primarily guided by the mean and the number of ratings, and -- to a lesser degree -- by the variance and origin of a rating. When probing the maximizing behavioral tendencies of our participants, other sensitivities regarding the summary of rating distributions became apparent. We thus instrumented a follow-up eye-tracking study to explore in more detail, how the choices of participants vary in terms of their decision making strategies.
Therefore, we also digged deeper into interpersonal differences of users when interacting with a platform offering recommendation services. We run a large user study (i.e. 300+ participants) simulating an online movie recommender and observed also respondents' browsing behaviour. Our results demonstrate that users manifesting maximizing behaviours also perform different search patterns, although their preferences appear to be the same.
As a conclusion, this thesis researches the relationship between ratings and recommendations from multiple perspectives and opens up several perspectives for research on explainable recommendations and their impact users' decision making.