Knowledge graph summarization impacts on movie recommendations

A classical problem that frequently compromises Recommender System (RS) accuracy is the sparsity of the data about the interactions of the users with the items to be recommended. The use of side information (e.g. movie domain information) from a Knowledge Graph (KG) has proven effective to circumvent this problem. However, KG growth in terms of size and complexity gives rise to many challenges, including the demand for high-cost algorithms to handle large amounts of partially irrelevant and noisy data. Meanwhile, though Graph Summarization (GS) has become popular to support tasks such as KG visualization and search, it is still relatively unexplored in the KG-based RS domain. In this work, we investigate the potential of GS as a preprocessing step to condense side information in a KG and consequently reduce computational costs of using this information. We propose a GS method that combines embedding based on latent semantics (ComplEx) with nodes clustering (K-Means) in single-view and multi-view approaches for KG summarization, i.e. which act on the whole KG at once or on a separated KG view at a time, respectively. Then, we evaluate the impacts of these alternative GS approaches on several state-of-the-art KG-based RSs, in experiments using the MovieLens 1M dataset and side information gathered from IMDb and DBpedia. Our experimental results show that KG summarization can speed up the recommendation process without significant changes in movie recommendation quality, which vary in accordance with the GS approach, the summarization ratio, and the recommendation method.