Dimension Reduction for Robust Covariate Shift Correction
share
In the covariate shift learning scenario, the training and test covariate distributions differ, so that a predictor's average loss over the training and test distributions also differ. The importance weighting approach handles this shift by minimizing an estimate of test loss over predictors, obtained via a weighted sum over training sample losses. However, as the dimension of the covariates increases, this test loss estimator increases in variance. In this work, we adapt the importance weighting approach to more robustly handle higher dimensional covariates by incorporating dimension reduction into the learning process.
READ FULL TEXT