局部子空间自适应核函数网络

针对目前单一整体核函数网络相对于原始空间特征局部分布缺乏自适应性以及网络规模巨大的不足,本文提出局部子空间自适应核函数网络解决思路。首先将原始空间的高维特征分成一系列低维子空间;然后针对不同子空间构造不同的核函数和核参数局部网络;最后综合所有子空间形成综合核函数网络。同时利用支持向量分布几何意义,选择边界向量代替原始训练样本,大大减少网络的复杂度。由于不同输入子空间构造的核函数以及网络规模相对于样本分布具有一定的自适应性,所以该方法相对于经典全局核函数网络可以得到更好的综合性能。对比实验结果表明本文提出的方法能够改进高维模式识别的性能,是一种值得实际应用推广的方法。

Local Subspace Adaptive Kernel Function Network

In order to overcome the deficiency of unitary kernel to global features and large-scale structure in general kernel func- tion network,a novel adaptive local subspace kernel function network is proposed in this paper.Firstly the input space is divided into a series of significant low dimensional subspace,and then the local adaptive kernel function and parameters are constructed in each sub- space.Finally the integrated kernel function network is setup by using all the information of each kernel subspace.At the same time the margin vector is selected based on a deep investigation into the geometry principle of support vectors and its distribution.Compared ex- perimental results show that this algorithm performs better in high dimensional pattern recognition application,may be applicable in many other fields.