A self-adaptive local metric learning method for classification

Mahsa Taheri, Zahra Moslehi, Abdolreza Mirzaei, Mehran Safayani

Abstract: Highlight•In this paper, a new framework named SA-LM2 in supervised local distance metric learning is introduced. In this framework, learning an appropriate distance metric and finding the local neighborhoods are integrated in a joint formulation.•Unlike other existing algorithms, where they need to select k nearest neighbors, SA-LM2 learns the radius of local neighbourhoods automatically.•SA-LM2 is expressed as a semidefinite programming, where due to its convex nature it avoids the local optima and is of global convergence guarantee.•Only the dissimilar set D is applied in SA-LM2, which makes it useful in some application where we do not have any prior knowledge about the similar data.•The results of SA-LM2 are less influenced by noisy input data points than the other compared global and local algorithms.