Manifold Regularized Transfer Distance Metric Learning

The performance of many computer vision and machine learning algorithms are heavily depend on the distance metric between samples. It is necessary to e xploit abundant of side information like pairwise constraints to learn a robust and reliable distance metric. While in real world application, large quantities of labeled data is unavailable due to the high labeling cost. Transfer distance metric learning (TDML) can be utilized to tackle this problem by leveraging different but certain related source tasks to learn a target metric. The recently proposed decomposition based TDML (DTDML) is superior to other TDML methods in that much fewer variables need to be learned. In spite of this success, the learning of the combination coefficients in DTDML still relies on the limited labeled data in the target task, and the large amounts of unlabeled data that are typically available are discarded. To utilize both the information contained in the source tasks, as well as the unlabeled data in the target task, we introduce manifold regularization in DTDML and develop the manifold regularized transfer distance metric learning (MTDML). In particular, the target metric in MTDML is learned to be close to an integration of the source metrics under the manifold regularization theme. That is, the target metric is smoothed along each data manifold that is approximated by all the labeled and unlabeled data in the target task and each source metric. In this way, more reliable target metric could be obtained given the limited labeled data in the target task. Extensive experiments on the NUS-WIDE and USPS dataset demonstrate the effectiveness of the proposed method.