Optimal Transport for Multi-source Domain Adaptation under Target Shift

  • 2019-03-14 15:38:56
  • Ievgen Redko, Nicolas Courty, RĂ©mi Flamary, Devis Tuia
  • 0

Abstract

In this paper, we propose to tackle the problem of reducing discrepanciesbetween multiple domains referred to as multi-source domain adaptation andconsider it under the target shift assumption: in all domains we aim to solve aclassification problem with the same output classes, but with labels'proportions differing across them. This problem, generally ignored in the vastmajority papers on domain adaptation papers, is nevertheless critical inreal-world applications, and we theoretically show its impact on the adaptationsuccess. To address this issue, we design a method based on optimal transport,a theory that has been successfully used to tackle adaptation problems inmachine learning. Our method performs multi-source adaptation and target shiftcorrection simultaneously by learning the class probabilities of the unlabeledtarget sample and the coupling allowing to align two (or more) probabilitydistributions. Experiments on both synthetic and real-world data related tosatellite image segmentation task show the superiority of the proposed methodover the state-of-the-art.

 

Introduction (beta)

None

 

Conclusion (beta)

None