一种局部优化边界的支持向量数据描述方法

Support vector data description method with local optimization boundary

针对传统的支持向量数据描述(SVDD)因未考虑数据构成的多模态性和局部分布的非同一性,难以获取目标数据的优化决策边界,所建立的数学模型难以正确反映建模对象的时空变化规律的问题,提出一种基于局部优化边界的支持向量数据描述(LOB-SVDD)方法。通过求取局部数据样本的分散程度获取支持向量机算法中折衷参数的局部调整系数,以此优化求解决策边界函数,由此可实现数据分类、离群点检测和数据建模等。利用UCI数据集和人工双模态数据集进行的仿真表明,与传统方法相比,LOB-SVDD可获得更优的决策边界,作为分类器有更低的假正率和假负率。应用LOB-SVDD对具有多模态特性的铜锍吹炼实际生产数据进行预处理,能有效检测离群点,剔除异常样本,实现数据洁净化。

Conventional support vector data description （ SVDD） , which did not consider multi-modal and local distribution difference of the data, failed to reflect time-space variety rule of the object and hard to gain the optimal decision boundary. To solve this difficulty, a new SVDD method with local optimization boundary （ LOB-SVDD） was proposed. First, the local dispersion degree of each data point was calculat-ed, then, the coefficient of trade-off parameters was adjusted with the local dispersion degree, finally, the quadratic programming problem was solved and an optimized boundary function was obtained. The method can be used in data classification, outlier detection and data modeling, etc. Experiments with UCI datasets and artificial dual mode datasets show that the method can gain a more optimal decision boundary compared to the conventional method, and as classifier it can gain lower false positives rate and false nega-tives rate. That method was applied to the multi-modal actual production data of copper matte converting process, and the results show that it can effectively detect outliers, eliminate abnormal sample data.