Short-term traffic prediction is one of the most important elements in Intelligent Transportation Systems (ITS). Although an extensive collection of state-of-the-art methods emerge, the well-established approaches require stringent assumptions that both the training data and testing data must be in the same feature space with independent and identical distribution and sufficient and representative training data, including most features in the feature space. However, due to systematic measurement error, communication problems, and device failure, traffic sensors cannot provide sufficient and erroneous-free data all the time in practice. To address this challenge, this paper proposes a feature-based domain adaptation framework for short-term traffic prediction. This proposed framework can obtain traffic flow feature representations across different domains (i.e., links), where the marginal probability distributions in different domains are close to each other, while the data structural properties in each domain are adequately preserved. With such a mapping, the cross-domain feature set can be created, and any conventional statistical learning methods can be applied to train regression models for the prediction task on the target links. The proposed domain adaptation framework is validated by using traffic flow data collected from Inductive Loop Detectors (ILD) in the UK. The results show that the proposed framework can transfer the features in the source domain to improve the prediction performance in the target prediction task.