采用划分融合双向控制的粒度支持向量机

Granular support vector machine with bidirectional control of division-fusion

粒度支持向量机(granular support vector machine,GSVM)引入粒计算的方式对原始数据集进行粒度划分以提高支持向量机(support vector machine,SVM)的学习效率。传统GSVM采用静态粒划分机制,即通过提取划分后数据簇中的代表信息进行模型训练,有效地提升了SVM的学习效率,但由于GSVM对信息无差别的粒度划分导致对距离超平面较近的强信息粒提取不足,距离超平面较远的弱信息粒被过多保留,影响了SVM的学习性能。针对这一问题,本文提出了采用划分融合双向控制的粒度支持向量机方法(division-fusion support vec-tor machine,DFSVM)。该方法通过动态数据划分融合的方式,选取超平面附近的强信息粒进行深层次的划分,同时将距离超平面较远的弱信息粒进行选择性融合,以动态地保持训练样本规模的稳定性。通过实验表明,采用划分融合的方法能够在保证模型训练精度的条件下显著提升SVM的学习效率。

Granular support vector machine(GSVM)introduces the method of granular computing to divide the original dataset;therefore,GSVM improves the efficiency of the support vector machine(SVM).The traditional GSVM adopts the static granules partitioning mechanism to extract representative information from the divided data clusters for model training,which can effectively increase the learning efficiency of the SVM.However,the GSVM uses the same pro-cessing way for different information granules,which may lead to a decline in the generalization ability because of two reasons:(i)No sufficient valid information is extracted from the strong information granules that are close to the hyper-plane,and(ii)excess of the weak information of granules far from the hyper-plane is reserved.These all reduce the learning performance of the SVM.To address this problem,this study proposes a division and fusion SVM model based on dynamical granulation,namely DFSVM.With the DFSVM,the information from the strong information granules near the hyper-plane is divided in depth,and weak information from weak information granules far from the hyper-plane is selectively merged to dynamically maintain the stability of the size of the training samples.The experiments demon-strate that this model can significantly improve the SVM learning efficiency,ensuring the training precision of the model.