Abstract:
In recent years, it has been theoretically verified that, compared with a single margin, margin distribution is more critical to the generalization performance. Although large margin distribution machine (LDM) can get superior classification and stronger generalization performance by maximizing the margin mean and minimizing the margin variance simultaneously, classifiers may be overwhelmed by the majority classes such that the minority class could have a lower detection rate due to ignoring the class imbalance. This is apparently contradict to the needs of high detection rate on the minority class in many real applications. Aiming at the above problem, this paper proposes an imbalanced cost-sensitive large margin distribution machine (ICS-LDM) to improve the detection rate of the minority class. First, when calculating the margin mean and margin variance, different weights are chosen on the sample margin between different types. And then, the objective function is optimized effectively by means of the cyclic dual coordinate descent method (Cyclic-DCD). Thus, a balanced distribution and maximum total margin is obtained by gradually increasing the margin distribution of the minority class. Finally, it is shown from experimental results that the proposed ICS-LDM can improve the classification accuracy of minority class and obtain more balanced detection rates on virtual dataset and UCI datasets.