Recommender systems are algorithms that are most well known for their applications in e-commerce. Given a specific customer and a large number of products, they automatically find the most relevant products to that specific customer. However, their relevance goes well beyond. They can also recommend genes responsible for diseases, words relevant to documents, tags relevant to a photo, courses of interest to a student etc.
The existing research on recommender systems is almost fully determined by the datasets that are (publicly) available. Therefore, the following fundamental question remains largely unstudied: "Given two datasets, how can we determine which of both has the highest quality for generating recommendations?"
Furthermore, the cornerstone of recommender systems research is the evaluation of the recommendations that are made by the recommender system. Most existing research relies upon historical datasets for assessing the quality of recommendations. There is however no convincing evidence that the performance of recommendations on historical datasets is a good proxy for their performance in real-life settings. Hence, also a second fundamental question remains largely unstudied: "How does the real-life performance of recommender systems correlate with measures that can be computed on historical data?"
By means of this project proposal, we set out to answer these two questions, which are foundational to recommender systems research.