Graph neural networks for recommendation leveraging multimodal information

n the era of digital information overload on the Internet, recommender systems act as filtering algorithms to provide users with items that might meet their interests according to expressed preferences and items’ properties and characteristics. Among the various recommendation paradigms so far, collaborative filtering has represented the most successful one thanks to the easy application of its recommendation algorithms whose high level performance in terms of recommendation accuracy is widely recognized. Despite the recent success of machine and deep learning techniques for collaborative filtering which have been effectively applied to recommender systems to improve the learning and profiling abilities of such models, recommendation still remains an highlychallenging task. Among the most debated open issues in the community, this thesis considers two algorithmic and conceptual ones, namely: (i) the inexplicable nature of users’ preferences, especially when they come in the form of implicit feedback; (ii) the effective exploitation of the collaborative information in the designing and training of recommendation models. In specific scenarios and domains such as fashion, food, tourism, and media content recommendation, the shallow item’s profile, commonly based upon the information conveyed within the user-item interactions only, may be enhanced through the multifaceted characteristics describing items. Driven from these assumptions, in the first part of this thesis, we propose to apply multimodal deep learning strategies for multimedia recommendation; the scope is to study and design recommendation algorithms based upon the principles of multimodality to possibly match each items’ characteristic to the implicit preference expressed by the user, thus addressing the (i) issue. Recent collaborative filtering approaches leverage the representational power of machine learning systems to profile users and items through embedding vectors in the latent space. In doing so, however, such recommendation models disregard a wide range of structural properties which are naturally encoded into the user-item interaction data. Indeed, recommendation datasets are easily describable under the topology of a bipartite and undirected graph, with users and items being the graph nodes connected at multiple distance hops. In this respect, the application of graph neural networks, recent machine learning techniques specifically tailored to learn from non-euclidean data, is of the utmost importance to provide a refined representation of users and items which can mine near- and long-distance relationships in the user-item graphs. Indeed, this is one possible way to exploit the collaborative signal, which is effectively propagated within the user-item graph, thus addressing the (ii) issue. Far from considering multimodal-aware and graph-based recommender systems from a separate perspective, this thesis conclusively aims to match the two families of recommendation strategies to propose an approach leveraging graph neural networks and multimodal information data. In seeking this joint research objective, other numerous micro aspects within the two macro areas (introduced above) are examined. Indeed, the thesis is a systematic compendium of careful additional analyses regarding, among the others, reproducibility, novel evaluation dimensions, and tasks and scenarios complementary to recommendation.