Algorithms increasingly influence the decisions people make in their everyday lives, from what clothing to buy to what movie to watch to what healthcare plan to adopt. There is a pressing need to understand how algorithms interact with human behavior in these settings. This project will develop theoretical foundations and applications for new machine learning algorithms that learn and predict human decisions descriptively from data, as they are, rather than as behavioral theories prescribe them to be. These new algorithms build on recent interpretations of choices as driven predominantly by pairwise interactions, involving new tools from graph theory to modeling human decision making. The unique practical potential of this project stems from its goal to operationalize these behavioral theories within a machine learning framework, making it possible to employ the lessons of behavioral economics to improve the design of large-scale web systems. The project builds on a set of key breakthroughs in the recent literature on machine learning for human decision making that model irrationality through interferences between alternatives in choice sets. The project aims to understand the theoretical limits of these interference-based approaches to modeling human choices, as well as to apply the methods to complex choice problems including ranking problems. The project also aims to adapt these new modeling tools to construct defensive tools capable of protecting people from malicious