最小二乘支持向量机分类器的高稀疏化及应用

为减少训练完毕之后的最小二乘支持向量机的分类计算量,借鉴神经网络的快速剪枝策略,提出了一种新的稀疏化算法:HS-LSSVM。它在主成分分析基础上,筛选出样本子集作为支持向量,它们既包含较多核函数矩阵信息,又相互独立性强,具有较好的代表性。算法将其余个体的信息转移至支持向量上,在实现高度稀疏化的同时,良好地保持了LSSVM的分类性能,并能适用于多类问题。对多个分类问题的测试表明,HS-LSS-VM具有稀疏率高,分类性能强,且稀疏化速度较快等优点。

High sparseness least squares support vector machine classifier

To decrease the classification computation load of a trained LSSVM,a new type of sparse algorithm for LSSVM,i.e.HS-LSSVM,is proposed referring to the neural networks pruning strategy.Based on the principle component analysis(PCA),a subset of training sample,whose corresponding kernel matrix columns contain most information of the kernel matrix and these columns are weakly linearly dependent to each other,is selected as the support vector(SV) set.The kernel matrix information contains in the non-SV will be transferred to the SV to hold good classification performance of LSSVM.HS-LSSVM can be used for multi classification.The experiments on several pattern classification problems show that HS-LSSVM has high sparseness while holds good classification performance and its sparse process is fast.