Robust domain-adaptive discriminant analysis

Consider a domain-adaptive supervised learning setting, where classifier learns from labeled data in source domain and unlabeled target to predict the corresponding labels. If classifier’s assumption on relationship between domains (e.g. covariate shift, common subspace, etc.) is valid, then it will usually outperform non-adaptive classifier. its invalid, can perform substantially worse. Validating assumptions relationships not possible without We argue that, order make classifiers more practical, necessary focus robustness; robust sense that an adaptive still at least as well having rely validity of strong assumptions. With this objective mind, we derive conservative parameter estimation technique, which transductive Vapnik Chervonenkis, show for discriminant analysis new estimator guaranteed achieve lower risk given samples compared Experiments problems with geographical sampling bias indicate our performs well.