针对传统支持向量机(SVM)增量算法,在学习过程中因基于局部最优解而可能舍弃含隐性信息的非支持向量样本,以及对于新增样本需全部进行训练的缺点,文中提出一种基于KKT条件和壳向量的SVM增量学习算法。该方法利用壳向量的特性保留了训练样本集中可能含隐性信息的非支持向量,并只将违反KKT条件的增量样本加入新的训练集,从而提高运算效率。通过对公共数据集Abalone和Balance Scale的实验表明,新算法在属性列数较多的数据集上分类效果更明显。
The traditional support vector machine (SVM) incremental algorithm in the learning process may give up the non-support vectors with implicit information, and requires the training of all the incremental samples. This paper presents a new incremental SVM learning algorithm based on KKT conditions and hull vectors. The algorithm makes use of the characteristics of hull vectors to retrain non-support vectors with implicit information, and it only add the samples violating the KKT conditions to the new training set. The experimental results from Abalone dataset and Balance Scale dataset show this algorithm has better classification effect in the datasets with more columns of properties.