LAM3L: Locally adaptive maximum margin metric learning for visual data classification

Visual data classification, which is aimed at determining a unique label for each class, is an increasingly important issue in the machine learning community. In recent years, increasing attention has been paid to the application of metric learning for classification, which has been proven to be a good way to obtain a promising performance. However, as a result of the limited training samples and data with complex distributions, the vast majority of these algorithms usually fail to perform well. This has motivated us to develop a novel locally adaptive maximum margin metric learning (LAM3L) algorithm in order to maximally separate similar and dissimilar classes, based on the changes between the distances before and after the maximum margin metric learning. The experimental results on two widely used UCI datasets and a real hyperspectral dataset demonstrate that the proposed method outperforms the state-of-the-art metric learning methods.