M.Sc. Ville Tanskanen defends his PhD thesis "Machine Learning for Human Behavior: Understanding and Imitating Choices" on Friday the 31st of October 2025 at 12 in the University of Helsinki Exactum building, Auditorium CK112 (Pietari Kalmin katu 5, basement). His opponent is Associate Professor Jussi P. P. Jokinen (University of Jyväskylä) and custos Professor Arto Klami (University of Helsinki). The defence will be held in English.
The thesis of Ville Tanskanen is a part of research done in the Department of Computer Science and in the Multi-source Probabilistic Inference group at the University of Helsinki. His supervisor has been Professor Arto Klami (University of Helsinki).
Machine Learning for Human Behavior: Understanding and Imitating Choices
Human behavior is a multifaceted phenomenon that spans many fields of science, each attempting to understand and model it from their own perspectives. In this work, we study human behavior from the perspective of machine learning: What can machine learning do for the benefit of other fields studying this phenomenon, and how to use machine learning for modeling human behavior.
Machine learning, and especially deep learning, offers ways to model arbitrarily complex phenomena accurately by learning to approximate them from data, in this case logged human behavior. High accuracy calls for complex models that often gain predictive power at the cost of interpretability, and require large amounts of data, which can be unattainable for human behavior. Simpler, discipline-specific models are less data-hungry and constructed from first principles, balancing the interpretability with limited predictive power.
We begin by focusing on interpretable models originating from behavioral economics, where human behavior is examined through the choices people make when faced with risky options. Specifically, these models consider problems where humans select between random variables, i.e. lotteries offering payouts with some probabilities, and are often studied in controlled experiments with participants answering hypothetical choice problems. We motivate that such lotteries are prevalent in the everyday life, and show how machine learning can transform the world into lottery representations required by the interpretable models. Later, we examine the issues that may arise when using these models outside of controlled experiments, where the lotteries that humans encounter may not provide rich enough information for accurate modeling.
We then shift our attention to the flexible models of machine learning, and consider two ways to model human behavior with them. First, we study human experts' optimization behavior by assuming that they follow a Bayesian optimization (BO) model. There, the traditional BO is inverted and instead of generating the optimization sequence, we observe an expert's sequence and infer their prior belief about the optimization problem. Second, we focus on modeling human behavior with reinforcement learning, viewing humans as reward-maximizing agents. We contribute to inverse reinforcement learning by proposing an efficient algorithm for learning reward preferences for a large population of users. The final focus of the thesis is on pure imitation of human behavior by learning human-like action policies from demonstrations. We study the popular approach of using foundation models in order to simplify the problem, and find that a task-shift, the discrepancy between foundation model's task and the imitation task, greatly affects the imitation performance.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available in the University of Helsinki open repository Helda at
Printed copies will be available on request from Ville Tanskanen: