Light PhD Seminar: Double-Weighting for Covariate Shift Adaptation in the Prognosis of COVID-19 infections

Title
Double-Weighting for Covariate Shift Adaptation in the Prognosis of COVID-19 infections

Abstract
Supervised learning can enable multiple important medical applications such as the prognosis of COVID-19 infections. These scenarios are often affected by a covariate shift, in which the marginal distributions of covariates of training and testing samples are different, but the label conditionals coincide. For instance, for the COVID-19 prognosis, predicting one wave using data collected in previous waves requires carrying out covariate shift adaptation that accounts for changes in the health data. The methods presented avoid the limitations of existing weighting methods for covariate shift adaptation by using a double weighting for both training and testing samples. The methods presented are based on minimax risk classifiers (MRCs) and utilize averages
of weighted training samples to estimate expectations at testing of weighted feature functions. We provide novel generalization bounds for our method that show a significant increase in the effective sample size compared with reweighted methods. The proposed method also achieves enhanced classification performance in experiments carried out with both synthetic and medical datasets.

This activity is funded by Axa Research Fund