Pseudo-Labeling for Covariate Shift Adaptation in Machine Learning

Watch a technical lecture from Columbia University researcher Kaizheng Wang exploring a novel covariate shift adaptation method using pseudo-labeling techniques. Learn about developing regression functions with minimal mean squared error across target distributions by leveraging both labeled and unlabeled data from different feature distributions. Understand the proposed approach of splitting labeled data to generate candidate models and create an imputation model for filling missing labels. Examine the bias-variance decomposition that reveals the importance of low-bias imputation models in this process. Discover how kernel ridge regression demonstrates the effectiveness of this adaptation method for handling unknown covariate shifts in machine learning applications.