Pursuing Privacy and Personalization in Recommender Systems: The Role of Federated Learning

In the era of big data, the massive data availability has created a tough world to navigate for users, who are often overwhelmed by alternatives of products, services, music, movies. Big data is also the fuel of machine learning applications and powerful tools like recommender systems that help users choose among the myriad of alternatives. Regrettably, this big data often contains sensitive information (e.g., socio-demographic attributes, social relations, browsing history, patterns of visited location or played music) that reflects their personal behavior. Although the performance of machine learning-powered services is strictly related to the amount of collected data, today, people pay more and more attention to their privacy, and international jurisdictions have legislated new laws to limit and control data collection. In this context, federated learning is a recent paradigm for performing on-device machine learning model training without requiring personal data collection. Today, federated learning is considered, together with privacy-preserving techniques like differential privacy and encryption, the best candidate to face the data privacy challenges in machine learning. This thesis summarizes our effort to bring the opportunities offered by federated learning in recommender systems. We propose a model that puts users in control of their data and grants high-quality recommendations even when users decide not to disclose part of their sensitive preferences. Moreover, we propose an entirely new federated recommender model based on a sparse entropy-weighted combination of feature embeddings. This work is motivated by our significant findings on the relation between federated data distribution and training efficiency and makes us realize that the on-device training of federated learning can lead not only to data privacy and control but also to highly personalized user-targeted recommender systems. All our findings are supported by extensive experiments and analyses on a wide range of recommendation dimensions. Finally, this dissertation is enriched by an extensive literature review on recommender systems, privacy-preserving paradigms and techniques, and a survey on the increasingly pivotal branch of privacy-preserving recommender systems.