Source Hypothesis Transfer for Zero-Shot Domain Adaptation

Making predictions in target unseen domains without training samples is frequent in real-world applications, such as new products' sales predictions. Zero-shot domain adaptation (ZSDA) has been studied to achieve this important but difficult task. An approach to ZSDA is to use multiple source domain data and domain attributes. Several recent domain adaptation studies have mentioned that source domain data are not often available due to privacy, technical, and contractual issues in practice. To address these issues, hypothesis transfer learning (HTL) has been gaining attention since it does not require access to source domain data. It has shown its effectiveness in supervised/unsupervised domain adaptation; however current HTL methods cannot be readily applied to ZSDA because we have no training data (even unlabeled data) for target domains. To solve this problem, we propose an HTL-based ZSDA method that connects multiple source hypotheses by domain attributes. Through theoretical analysis, we derive the convergence rate of the estimation error of our proposed method. Finally, we numerically demonstrate the effectiveness of our proposed HTL-based ZSDA method.