Bayesian multi-source domain adaptation

Based on the Bayesian learning principle (BayesMSDA), this paper presents a new multi-source domain adaptation framework, where one target domain and more than one source domains are used. In this framework, the label of a target data point is determined according to its posterior probability, which is calculated using the Bayesian formula. To fulfill this framework, a novel prior of the target domain based on Laplacian matrix and a new likelihood that is dynamically obtained using the k-nearest neighbors of a data point are defined. We focus on the situation that there are no labeled data obtained from the target domain while most of them are from source domains. Experiments on synthetic data and real-world data illustrate that our framework has a good performance.