A Causal Lens on Facial Micro-Expressions and Components: Recognition, Robustness, and Synthesis

This thesis investigates how to make computer systems more accurate and fair in recognising human emotions from subtle facial movements called micro-expressions. It identifies hidden biases in current machine learning models and proposes new methods to reduce them, improving reliability and fairness across different people. The research also develops a technique to protect emotional privacy and creates a large, realistic video dataset to train better models. These advances help build safer and more trustworthy emotion-recognition technologies for real-world use in health, security, and human–computer interaction.