The Taub Faculty of Computer Science Events and Talks

Strategic classification studies learning in settings where agents can modify their features to obtain favourable outcome. Most current works focus on simple decision rules that trigger independent agent responses. Here we examine the implications of learning with more elaborate models that break the independence assumption. We present two works, each studying different (but related) models and tasks, and in which dependencies are introduces through space (using graphs) and time. Our first work considers learning with graph neural networks (GNNs) – neural architectures which make use of social relations between users to improve classification. We show how relying on the graph inadvertently introduces inter-user dependencies that significantly affect predictive outcomes. Our second work focuses on recommendation, and studies how learning personalized item scores incentivizes exposure-maximizing content creators to update their items. We argue that this incentive structure can essentially eliminate diversity in recommendations, but at the same time, can be used to promote intrinsic diversity over time – a process in which the underlying user-item graph is key. In both works, we use analysis and simulative experiments to show how strategic responses can either work against the system – or in its favor, and propose differential frameworks for learning in ways that promote both system and societal interests.