This PhD project builds on the group's existing work on stable and causal feature learning, explainability, and bias detection and mitigation. Its overarching objective is to develop learning methods that are more robust to distribution shifts and non-stationary environments while retaining interpretable and causally meaningful representations of such changes.
The thesis is motivated by the following research question: Can a challenging high-dimensional density-ratio estimation problem be decomposed into simpler, causally meaningful ratios and estimated along diffusion-style probability paths, yielding estimates that are sufficiently accurate for adaptation while also providing structured explanations of why two distributions differ?
Although recent work has explored path-based approaches to density-ratio estimation, important questions remain regarding their integration with structured representations and causal models of distribution shift. The project will therefore investigate how probability-path-based estimation can be combined with causal mechanism factorization, task-sufficient and stable representations, calibrated uncertainty quantification, and interpretable explanations of distribution shifts.
The expected contribution goes beyond the development of another deep density-ratio estimation method. Instead, the thesis aims to establish a unified framework that combines probability paths with causal mechanism factorization, stable task-relevant representations, and principled decompositions of distribution shift. By explicitly linking density-ratio estimation to the mechanisms underlying distributional changes, the project seeks to address a distinct research gap in the rapidly developing literature, which has so far focused predominantly on the construction and estimation of probability paths themselves.
