Mutual Domain Adaptation

Abstract

To solve the label sparsity problem, domain adaptation has been well-established, suggesting various methods such as finding a common feature space of different domains using projection matrices or neural networks. Despite recent advances, domain adaptation is still limited and is not yet practical. The most pronouncing problem is that the existing approaches assume source-target relationship between domains, which implies one domain supplies label information to another domain. However, the amount of label is only marginal in real-world domains, so it is unrealistic to find source domains having sufficient labels. Motivated by this, we propose a method that allows domains to mutually share label information. The proposed method finds a projection matrix that matches the respective distributions of different domains, preserves their respective geometries, and aligns their respective class boundaries. The experiments on benchmark datasets show that the proposed method outperforms relevant baselines. In particular, the results on varying proportions of labels present that the fewer labels the better improvement.