Simpler large margin distribution machine via weighted linear loss for large-scale classification

A large margin distribution machine (LDM) with margin distribution optimization guarantees the good generalization performance of the model. However, the existing LDM model uses a hinge loss, which suffers from low computational efficiency. To solve this problem, we propose a simpler weighted linear loss LDM model (SWLLDM). Our SWLLDM is based on weighted linear loss and LDM, but it is not a simple combination. On the one hand, our SWLLDM has the margin distribution optimization, which leads to better generalization performance. In SWLLDM, the margin variance terms are reduced and the margin mean is eliminated. On the other hand, our SWLLDM has better computational efficiency. In SWLLDM, the inequality constraint is removed. All samples are used to optimize the QPP with equality constraints. Finally, we perform a series of numerical experiments on UCI benchmark datasets, NDC datasets and steel surface defect dataset. The final results illustrate the feasibility and effectiveness of the SWLLDM algorithm.