Semi-supervised Coefficient-Based Distance Metric Learning

Distance metric learning plays an important role in real-world applications, such as image classification and clustering. Previous works mainly learn a distance metric through learning a Mahalanobis metric or learning a linear transformation. In this paper, we propose to learn a distance metric from a new perspective. We first randomly generate a set of base vectors and then learn a linear combination of these vectors to approximate the target metric. Compared with previous distance metric learning methods, we only need to learn the coefficients of these base vectors instead of learning the target metric or the linear transformation. Consequently, the number of variables needed to be determined is the same as the number of base vectors, which is irrelevant to the dimension of the data. Furthermore, considering the situation that labeled samples are insufficient in some cases, we extend our proposed distance metric learning method into a semi-supervised learning framework. Additionally, an optimization algorithm is proposed to accelerate training of our proposed methods. Experiments are conducted on several datasets and the results demonstrate the effectiveness of our proposed methods.