Multi-source Domain Adaptation via Weighted Joint Distributions Optimal Transport

The problem of domain adaptation on an unlabeled target dataset using\nknowledge from multiple labelled source datasets is becoming increasingly\nimportant. A key challenge is to design an approach that overcomes the\ncovariate and target shift both among the sources, and between the source and\ntarget domains. In this paper, we address this problem from a new perspective:\ninstead of looking for a latent representation invariant between source and\ntarget domains, we exploit the diversity of source distributions by tuning\ntheir weights to the target task at hand. Our method, named Weighted Joint\nDistribution Optimal Transport (WJDOT), aims at finding simultaneously an\nOptimal Transport-based alignment between the source and target distributions\nand a re-weighting of the sources distributions. We discuss the theoretical\naspects of the method and propose a conceptually simple algorithm. Numerical\nexperiments indicate that the proposed method achieves state-of-the-art\nperformance on simulated and real-life datasets.\n