Algorithmic Pursuit of Causality

Yihong Gu, PhD Candidate in Operations Research and Financial Engineering at Princeton University

The past few decades have witnessed remarkable advances in modern machine learning, particularly in deep learning and large language models, who now lead the state-of-the-art performance of the powerful prediction systems. However, from a statistical view, these methods often inherit a fundamental limitation: the learning target is to find the most predictive solution in population, which inevitably reflects intrinsic biases present in the collected data. This limitation results in fitted models that misrepresent causal relationships and so hinders the further intelligence of the system.

In this talk, I will present a solution to this challenge, particularly when domain knowledge is unavailable. The refined objective is designed to find a set of variables that yield invariant predictions across diverse environments while minimizing prediction error. I will discuss the proposed scalable and sample-efficient estimation framework, the causal interpretation of the refined target, the fundamental computational limits in theory, and practical approaches for efficient computation.