A reproducibility analysis of privacy-oriented approaches for Federated Learning to Rank

The raise of Artificial Intelligence and Machine Learning techniques together with the rapid growth of Information Retrieval systems, emphasize the fundamental role of the Learning to Rank (LtR) models. LtR models provide relevant and personalized results to the users. Traditional LtR approaches typically rely on a centralized data collection, which can present some privacy concerns, when sensitive data and user interactions are involved. To mitigate these challenges, Federated Learning (FL) has been proposed, where in this framework the user private data is not shared to a centralized server, but remains local and only device updates are shared. In this context, Federated Learning to Rank (FLtR) integrates the ranking models within the FL framework, combining the advantages of a distributed paradigm with the ranking objectives. This work presents the analysis of some existing privacy-preserving approaches of FLtR, with a focus on their design principles, the optimization strategies involved, and the privacy-preserving mechanisms integrated. We design and implement a unified and modular framework that supports the reproducibility and evaluation of different FLtR methods, facilitating systematic comparison. With this study, the purpose of the thesis is to identify the trade-offs among reproducibility, privacy preservation and accuracy in FLtR, outlining potential directions for future research.