Thesis title: Decision Support Mechanisms for Socially Sensitive Contexts
In this PhD thesis, I will delve into the challenges of designing decision support mechanisms in settings where social and individual impact significantly shape and restrict the space of the available design options. Over the past four years, my research has specifically focused on developing algorithmically fair college admission criteria for disadvantaged minorities, on identifying misinformation when the labeling budget is limited, and on training virtual assistants to handle infrequent requests. Although these topics may appear disparate, they all involve practical, real-world scenarios where standard approaches have proven ineffective. In the first case study, published at ACM SAC, I propose an affirmative action policy for selecting university candidates that balances the representation of disadvantaged groups - while still selecting candidates who are more likely to perform well. In the second case study, presented in two separate papers - one published at IJCNN and the other currently under review, I explore the problem of selecting and labeling the right pieces of news to train state-of-the-art misinformation detection methods. After collecting and processing a new multilingual and challenging fake news benchmark dataset, I designed a deep active learning procedure to train graph-neural-networks-based fake news detectors under labeling budget constraints. Finally, in my third and last contribution, I develop a new clustering-aware loss for training sentence encoders. Resulting embeddings can be effectively used to detect novel and rare intents in the pool of unhandled requests of a voice assistant. This last piece of research has been conducted while I was interning at Amazon Alexa AI and is currently under review. Overall, this PhD thesis presents a thorough examination of the specific challenges encountered in designing decision support mechanisms in realistic and socially-conscious settings in which a strong imbalance of some sort exists. I hope my work will demonstrate that by considering the real-world social and user impact of designed solutions, it is possible to create more effective and robust automated decision-support systems.