NATIONAL PHD IN ARTIFICIAL INTELLIGENCE

Thesis title: Designing Secure and Knowledge-Aware Recommender Systems Leveraging Data Properties and Graph Structures

In an era where digital platforms inundate users with vast options, recommender systems (RSs) have become essential tools for filtering and personalizing content. Traditional approaches like collaborative filtering (CF) have limitations in handling complex data structures and ensuring user privacy and system security. This thesis addresses these challenges by designing secure and knowledge-aware recommender systems that leverage data properties and graph structures. We start by interpreting recommendation data as a graph, specifically a user-item bipartite graph, and explore how Graph Neural Networks (GNNs) can effectively model these structures to enhance recommendation accuracy. By analyzing the relationship between graph structures and GNN-based recommenders, we propose methods that optimize the utilization of inherent data properties, leading to more precise and relevant recommendations. Building upon the graph-based foundation, we integrate Knowledge Graphs (KGs) to enrich the semantic information within the recommendation data. By connecting items and users to entities in KGs, we augment the user-item graph with additional layers of meaningful relationships. We develop knowledge-aware recommendation models that combine the strengths of CF and GNNs, demonstrating that the incorporation of rich semantic data significantly improves recommendation performance. Addressing the critical issues of data privacy and security, we apply Differential Privacy techniques to the realm of recommender systems. We propose privacy-preserving strategies for the secure exchange and collection of recommendation data, ensuring compliance with regulatory frameworks like GDPR and enhancing user trust by safeguarding sensitive information. Furthermore, we tackle the vulnerability of recommender systems to adversarial attacks, where malicious entities manipulate data to influence recommendations unfairly. We introduce adversarial training methods to bolster the robustness of visual recommender systems against such attacks, thereby maintaining the integrity and reliability of the recommendations provided. Throughout this thesis, we present comprehensive analyses, novel methodologies, and practical tools to advance the field of recommender systems. By strategically leveraging data properties and graph structures, our contributions not only enhance recommendation effectiveness but also address vital security and privacy concerns, paving the way for more trustworthy and intelligent recommender systems.