Steffen Bickel, Christoph Sawade, & Tobias Scheffer Transfer Learning by Distribution Matching for Targeted Advertising
 Multiple web portals, biased samples  transfer learning across portals.  Web pages accessed by users, click stream  input to predictor. Portal 1 Portal 2 (auxiliary data)

Data for each portal:  unlabeled, governed by test distribution.  labeled, collected through questionnaires  selection bias.
Age: 28, Interests: a Fitness, Arts, Fashion Age: 49, Interests: a Cars, Finance, Golf Age: 35, Interests: a Fashion, Wine, Food

Transfer Learning by Distribution Matching: Pooled training distribution over all portals is matched to test distribution using resampling weights:

Discriminative reformulation of resampling weights  estimation with logistic regression.