CAJ | 학술논문

为了解决大规模非线性分类中局部学习的不平衡性问题，提出一种改进的局部支持向量机算法，在高维特征空间中聚类后，为每一个簇构造局部非线性支持向量机。为了克服簇内样本的分布不均衡问题，根据闭合超平面不规则边界的几何特点，经过梯度下降寻找稳定均衡向量，以此构造簇几何中心；再结合簇密度中心共同约束类心形成双重加权中心。然后通过求解加权最小闭球问题实现对大规模样本向量的分类。对照实验显示，除了个别数据集以外，改进的算法在训练时间、测试时间以及测试精度等方面都比另外两种分类算法表现更佳。
위료해결대규모비선성분류중국부학습적불평형성문제，제출일충개진적국부지지향량궤산법，재고유특정공간중취류후，위매일개족구조국부비선성지지향량궤。위료극복족내양본적분포불균형문제，근거폐합초평면불규칙변계적궤하특점，경과제도하강심조은정균형향량，이차구조족궤하중심；재결합족밀도중심공동약속류심형성쌍중가권중심。연후통과구해가권최소폐구문제실현대대규모양본향량적분류。대조실험현시，제료개별수거집이외，개진적산법재훈련시간、측시시간이급측시정도등방면도비령외량충분류산법표현경가。
An improved algorithm for localized support vector machine is proposed to resolve the imbalance of local learning problem in nonlinear classifications on large data sets. The algorithm uses the supervised clustering algorithm for clustering in a feature space of high dimension and then constructs local nonlinear support vector machines for each cluster. According to the geometric feature of irregular borders of enclosing sphere, the geometric center for a stable equilibrium point is constructed and a dual-weighted center of two relevant weights is formed through calculating density center of the cluster. At last, the classification of large data set is carried out by solving the problem of weighted minimum enclosing ball. Compared with the other two algorithms of controlled group, the proposed algorithm shows shorter training time and testing time as well as higher testing precision except for some individual data sets.