Online Low-Rank Metric Learning via Parallel Coordinate Descent Method

摘要: 11The corresponding author is Prof. Yang Cong. This work supported by Nature Science Foundation of China under Grant (61722311, U1613214, 61533015) and CAS-Youth Innovation Promotion Association Scholarship (2012163)Recently, many machine learning problems rely on a valuable tool: metric learning. However, in applications, large-scale applications embedded high-dimensional feature space may induce both computation storage requirements to grow quadratically. In order tackle these challenges, this paper, we intend establish robust formulation with the expectation that online parallel optimization can solve data efficiently, respectively. Specifically, based matrix factorization strategy, first step aims learn similarity function objective for measurement; second step, derive variational trace norm promote low-rankness transformation matrix. After converting regularization into its separable form, model optimization, present an block coordinate descent method optimal parameters, which handle efficient way. Crucially, our shares efficiency flexibility method, it also guaranteed converge solution. Finally, evaluate approach analyzing scene categorization dataset tens thousands dimensions, experimental results show effectiveness proposed model.