Slides: link
Abstract In an era where recommender systems wield significant influence over users' choices and experiences, prioritizing fairness and explainability is becoming imperative to ensure a responsible adoption. However, these two dimensions have been often considered independently from each other, which has resulted in missed opportunities for a mutual improvement. For instance, applying explainability techniques can lead to valuable insights into how fairness is achieved or compromised in the recommendation process and, therefore, better inform the design of algorithmic countermeasures against the identified discriminatory outcomes. On the other hand, recommendation methods able to explain why an item is suggested to the user may inadvertently reinforce existing biases or disadvantage certain groups of users, also by systematically providing them with explanations of lower quality. This talk will raise attention to the importance of combining fairness and explainability principles in order to develop more responsible recommendation processes. It will present motivating examples and emerging algorithmic solutions that showcase the benefit of integrating and assessing both these beyond-utility dimensions, pointing to recent studies that are based on public source code and datasets. This talk will finally summarize the open issues and the future directions this vibrant field is headed, encouraging a shift towards integrating fairness and explainability principles from the outset.
Bio Mirko Marras is Assistant Professor at the Department of Mathematics and Computer Science of the University of Cagliari (Italy), where he co-leads the research unit on responsible machine learning. Prior to that, he has been postdoctoral researcher at EPFL (Switzerland) and visiting scholar in numerous international institutions, such as Eurecat Technology Center (Spain), University of Las Palmas (Spain), and New York University (USA). He has co-authored more than 60 papers in top-tier conferences (such as UMAP 2023, SIGIR 2022, CIKM 2021) and journals (such as Elsevier's Information Processing & Management and Springer's User-Modeling and User-Adaptive Interaction). His research ranges across a wide range of domains impacted by user modeling and personalization, including business, education, entertainment, and healthcare. He is regularly part of the program committee of top conferences in the crossing fields (such as ECIR, RecSys, SIGIR, UMAP, WSDM, WWW). He has given tutorials on this theme at RecSys 2022, ICDE 2021, ECIR 2021, WSDM 2021, ICDM 2020, and UMAP 2020. He has also co-chaired several workshops, including the regular series on Bias at ECIR (2020-2023), the L2D workshop at WSDM 2021, the R&PRMI workshop at ICCV 2021, the FATED workshop at EDM 2022, and the RKDE workshop at ECML-PKDD 2023. He serves as associate editor for the Springer's Journal of Ambient Intelligence and Humanized Computing.