Interpretable domain adaptation using unsupervised feature selection on pre-trained source models

We study a realistic domain adaptation setting where one has access to an already existing “black-box” machine learning model. Indeed, in real-life scenarios, efficient pre-trained source predictive model is often available and required be preserved. The solution we propose this problem the asset of providing interpretable target transformation by seeking sparse ordered coordinate-wise feature space addition elementary mapping functions. To automatically select subset features adapted, first introduce weakly-supervised process relying on scarce labeled data. Then, address more challenging unsupervised version scenario. end, new pseudo-label estimator over unlabeled examples, which based rank-stability regards prediction. Such estimated “labels” are further used selection assess whether each needs transformed achieve adaptation. provide theoretical foundations our method as well implementation. Numerical experiments real datasets show particularly encouraging results since approaching supervised case, samples.