Evaluation of Explanations Generated by Knowledge-aware Recommender Systems with regard to Consumer and Provider Fairness

Knowledge-aware recommender system incorporate external information to improve recommendation performance, especially when user-item interactions are sparse. A key advantage of these approaches is that they can produce explanations for recommendations, which increases amongst others trust, transparency and interpretability of the models. Despite the increasing interest in explanation evaluation and fairness in recommmender systems research, the fairness of explanations has received little attention. This thesis adresses this gap by proposing a novel evaluation framework, which introduces two metrics measuring consumer-side as well as provider-side fairness. First, the interaction item bias disparity examines how well users’ category preferences are reflected by the models in the explanations. Second, the shared entity visibility and exposure quantifies how fairly different demographic groups of knowledge graph entities are represented in the explanations. The results indicate a systematic discrimination against minority groups, which increases with higher group imbalance.