A theory of learning from different domains

This paper studies the theoretical foundations of learning from different domains, a setting where plentiful labeled data exists in a source domain but the goal is to classify well in a target domain that has a different distribution and little or no labeled data. It addresses two questions: under what conditions a source-trained classifier can be expected to perform well on target data, and how to combine a small amount of labeled target data with abundant source data during training. The analysis bounds the target error using the source error together with a classifier-induced divergence measure that can be estimated from finite, unlabeled samples of the two domains.

Under the assumption that some hypothesis performs well in both domains, the source error and this divergence characterize the target error of a source-trained classifier. The authors further bound the target error of a model that minimizes a convex combination of the empirical source and target errors, and show how to choose the optimal combination as a function of the divergence, the sample sizes, and the hypothesis class complexity. The resulting bound generalizes previously studied cases and is always at least as tight as bounds that consider only the target error or an equal weighting of the two.