Deep Models of Decision

Decision Theory is a branch of mathematics and economics that studies the process of decision-making under uncertainty. It provides a framework for analyzing the choices made by individuals or organizations and gives guidance on how to make the best decisions possible given the limited information and resources available. Choice Modeling is a subfield of Decision Theory that aims to explain and measure how people or groups make decisions when presented with various options. It achieves this by modeling the link between the characteristics of those options and the likelihood that a person will select one over another. Some of its applications include analyzing consumer behavior in marketing research, predicting voter preferences in political campaigns, optimizing public policy decisions, and designing recommendation systems for online platforms. The main goal of this thesis is to propose and assess novel approaches to Choice Modeling that leverage the expressivity of deep machine learning models. The first approach we investigate is an application of deep learning in conjunction with Quantum Decision Theory, a framework developed by cognitive psychologists that proposes a way of modeling phenomena such as uncertainty and interactions between alternatives inspired by quantum mechanics. The second one is based on the Attention Mechanism, notoriously used in deep learning for modeling long-distance relationships between inputs. We aim to explore the capabilities of these new models and provide a general way of applying them to new choice problems. We evaluate the efficacy of these new methods on three different Choice Modeling datasets of increasing complexity. Furthermore, we compare our methods against reference models from both Classical and Quantum Decision Theory. Finally, we discuss potential avenues for future research.