一种改进的基于支持向量机的多类分类方法

摘要

针对现有支持向量机多类分类算法在分类精度上的不足，提出一种改进的支持向量机决策树多类分类算法。为了最大限度地减少误差积累的影响，该算法利用投影向量的思想作为衡量类分离性的标准，由此构建非平衡决策树，并且在决策树节点处对正负样本选取不同的惩罚因子来处理不平衡数据集的影响，最后引入KNN算法与SVM共同识别数据集。通过在手写体数字识别数据集上的仿真实验，分析比较各种方法，表明该方法能有效提高分类精度。%In light of the deficiency of existing SVM multi-class classification algorithm in classification accuracy, we propose an improved SVM decision tree multi-class classification algorithm.In order to minimise the impact of the error accumulation to greatest extent, the algorithm uses the idea of vector projection as the standard to measure class separation, thus constructs an unbalanced decision tree.Furthermore, it selects different punishment factors from positive and negative samples at the nodes of decision tree to counteract the impact from unbalanced data sets.At last, it introduces KNN to co-recognise the data sets with SVM.Analysing and comparing diffident methods by the simulation experiment on handwritten digit recognition data sets, it is shown that this method can effectively improve the classification accuracy.