单双点平滑结合的流形正则化半监督分类学习框架

SEMI-SUPERVISED LEARNING FRAMEWORK FOR MANIFOLD REGULARIZATION BASED ON THE SMOOTH COMBINATION OF SINGLE AND DOUBLE POINTS

提出一种单双点平滑结合的流形正则化半监督分类学习框架(SDS_MR)。流形正则化(Manifold regularization,MR)约束流形图上的相似实例应该获得相似的分类输出,但是MR是建立在流形图上的双点平滑上的,即平滑约束是在所有实例对上实现的,也就是视每个实例对为单位。在MR的基础上做了改进,通过结合实例对平滑和单个实例的平滑来实现半监督学习。通过这种方式,单双点的平滑都被取利用,都具重要性且做出贡献,这种重要性可以是双点约束信息,也可以是单点局部密度。在UCI标准测试数据集上的实验结果显示,SDS_MR与MR相比具有一定的竞争力。

A new semi-supervised learning framework for manifold regularization based on the smooth combination of single and double points(SDS-MR)is proposed.MR(Manifold Regularization)constrained similar instances on manifold graphs to obtain similar classification output.But MR was based on the double points smoothing of manifold graph,the smoothing constraint was implemented on all instance pairs,that is,each instance pair was regarded as a unit.On the basis of MR,SDS_MR combined the smoothing of an instance with the smoothing of a single instance.In this way,the smoothing of single and double points was used,both of which were important and contribute.This importance could be either double points constraint information or single point local density.The empirical results on UCI standard test data set show that SDS_MR is more competitive than MR.