A framework for learning and analyzing hybrid recommenders based on heterogeneous semantic data

With the constantly growing amount of online available products and data, recommender algorithms are one of the key technologies helping users to cope with information overload. In contrast to information retrieval systems, recommender systems discover new items matching users' preferences. An item's relevance depends on a collection of criteria, such as its properties, user preferences and context. In this paper, we focus on recommending previously unrated items. We analyze how to compute highly relevant recommendations based on aggregated content-based knowledge. We compare different approaches for aggregating semantic knowledge and show how to find good scaling and recommender models taking into account the specific dataset properties. Furthermore, we discuss the gain of combining different semantic data sets. We evaluate our approaches on semantic movie datasets as well as on user feedback collected with our movie recommender web application. The evaluation shows that for each semantic relationship set a specific recommender model should be learned. Learning a "global" recommender for the aggregated dataset (consisting of several heterogeneous datasets) results in a lower recommendations quality than creating an ensemble of recommenders. We study the complete recommendation creation process from the semantic dataset to a web application discussing the challenges and pitfalls of each step.