Tommaso Carraro

      Ritratto di Tommaso Carraro

Computer Science and Innovation for Societal Challenges, XXXVII series 

Grant sponsor

Luciano Serafini & Fabio Aiolli 



Project description

Recommender system (RS) technologies are nowadays an essential component in commercial applications (e.g., Amazon, Netflix, Spotify). Due to the volume and variety of items offered by these platforms, it is imperative to have a tool that can help users in selecting new items according to their preferences. Specifically, the goal of an RS is to suggest items (e.g., movies, songs, news) that are most likely of interest to a particular user. The story of RSs began in early 2000, with the release of methods based on Latent Factor Models. In particular, Matrix and Tensor Factorization were the most prominent ones. Over the years, as in most applications, deep learning has shown of outperforming existing methods, becoming very soon the standard approach to tackle the recommendation problem. However, despite their success, state-of-the-art deep approaches still suffer from important limitations. On the one hand, researchers are still trying to deal with sparsity and cold-start problems, on the other hand, the black-box nature of these models often creates skepticism among users. To face these problems, researchers' attention moved towards Knowledge-based RSs. Research in this field has proven that the injection of background knowledge is crucial to deal with sparsity and explainability issues. Among the knowledge-based approaches, Neural-Symbolic Integration has shown to be effective in many AI fields. This branch of Artificial Intelligence aims at integrating deep learning with logical representation and reasoning. The idea is to merge the advantages of the two paradigms in a unified framework. Deep learning is well known for its robustness to noise, while logical reasoning is interpretable by design and allows learning from a few samples. My Ph.D. will be focused on experimenting with the application of Neural-Symbolic approaches to recommender systems in order to address their limitations. This research area is still in its initial stages and raises multiple challenges, such as the choice of the right logical language, the generation of a reliable knowledge base, and the study of methods to perform an effective and efficient integration of the paradigms.