[NOVA Math] Seminar of Data Science

The Center of Mathematics and Applications (NOVA Math), promote the Seminar of Data Science with the title: “A Theoretical and Robustness Analysis of Recommender Systems”.  Giulia Di Teodoro (University of Pisa) and Federico Siciliano (Sapienza University of Rome) are the speakers.

Abstract: Recommender Systems (RSs) are pivotal in diverse domains such as e-commerce, music streaming, and social media.

The first part of the seminar presents a comparative analysis of key loss functions in recommender systems: Binary Cross-Entropy (BCE), Categorical Cross-Entropy (CCE), and Bayesian Personalized Ranking (BPR), which distinguish between positive items (interacted by users) and negative items. While previous studies have empirically shown that CCE outperforms BCE and BPR with the full set of negative items, we provide a theoretical explanation by proving that CCE offers the tightest lower bound on ranking metrics like Normalized Discounted Cumulative Gain (NDCG). Given that using the full set of negatives is computationally expensive, we derive bounds for these losses in negative sampling settings, establishing a probabilistic lower bound for NDCG. Our analysis shows that BPR's bound on NDCG is weaker than BCE’s, challenging the common belief that BPR is superior to BCE in recommender system training.

Beyond loss function analysis, we turn our attention to the robustness of Sequential Recommender Systems against dataperturbations. Traditional similarity measures, such as Rank-Biased Overlap, prove inadequate for evaluating ranking stability in finite-length sequences. To address this, we introduce Finite Rank-Biased Overlap, a novel similarity measure tailored for practical scenarios. Through empirical analysis of item removal at different positions in temporally ordered sequences, we demonstrate that the impact on recommendation quality is highly position-dependent, with removals at the end of sequences leading to significant performance degradation.

Giulia Di Teodoro is currently a postdoctoral researcher at the University of Pisa, specializing in machine learning and information-filtering systems. She graduated with honors in Management engineering from Sapienza University of Rome, Italy, where she also completed her PhD in Data Science in 2024, with a thesis on precision medicine for HIV and diabetes and interpretability of machine learning models. She has publications in the domain of Precision Medicine, Explainable Artificial Intelligence and Bioinformatics. Her academic journey includes extensive international exposure, such as a visiting researcher position at the Uniklinik Köln in Germany, in collaboration with the  Max Planck Institutes. Her research interests encompass Mixed-Integer Linear Programming (MILP), Precision medicine and Recommendation Systems. 

Federico Siciliano is a postdoctoral researcher at Sapienza University of Rome, specializing in information retrieval and recommender systems. He completed his PhD in Data Science at Sapienza University in 2024, with a thesis on the architectural components of trustworthy artificial intelligence, earning the distinction of "excellent with honors".

Federico's professional experience includes a research internship at Amazon Italy. He has also contributed to various research grants.

As an active member of the academic community, Federico has served as a PC member for various conferences and organized workshops at both SIGIR and RecSys. His expertise spans recommender systems, robust AI, and trustworthy models, evidenced by publications in top-tier venues such as RecSys, SIGIR, and IJCNN.

March 26, 2025, 14:00, Library Auditorium, NOVA FCT.