Learning neighborhoods for metric learning

摘要: Metric learning methods have been shown to perform well on different tasks. Many of them rely target neighborhood relationships that are computed in the original feature space and remain fixed throughout learning. As a result, learned metric reflects relations. We propose novel formulation problem which, addition metric, relations also two-step iterative approach. The new can be seen as generalization many existing methods. includes neighbor assignment rule assigns numbers neighbors instances according their quality; 'high quality' get more neighbors. experiment with two its instantiations correspond algorithms LMNN MCML compare it other number datasets. experimental results show state-of-the-art performance provide evidence does improve predictive performance.