Optimizing margin distribution for online multi-label classification

This paper introduces a novel online multi-label classification approach that focuses on optimizing the distribution of multi-label classification margins. The objective is to find a new multi-label model in each online round that maximizes the margin mean and minimizes the margin standard deviation while remaining close to the current model. To address this optimization problem, two methods are proposed. The first method employs dual coordinate descent to solve the dual problem iteratively, while the second method utilizes online gradient descent to solve the first-order approximation of the primal problem efficiently. Experimental evaluations are conducted on nine benchmark datasets, using seven common multi-label performance metrics. The results demonstrate the superiority of our proposed approach over existing state-of-the-art methods, affirming the effectiveness and efficiency of optimizing the distribution of multi-label classification margins.