LPA-SKFST半监督特征提取方法

针对传统LDA类半监督特征提取方法的解矢量非正交、解空间不稳定和非线性处理能力不足等问题,提出LPA-SKFST方法.该方法的前置级LPA通过标签传播提高标记样本容量,后置级SKFST(半监督核最佳鉴别矢量集)采用双向正则方法对KFST引入全局结构保持正则和Tikhonov正则,并以成对空间求解方法求取Fisher分母矩阵奇异和非奇异时的统一形式解.在circle、iris、wine和自有珍珠光谱集的分类实验中,PCA、LDA、SLDA和SDG组的准确率随样本集、标记样本占比和标签可靠性变化而波动,LPA-SKFST组则稳定保持在85％以上.该结果证明,LPA-SKFST能克服标记样本占比和标记可靠性不足局限,在实际集和线性不可分人工集上取得一致、稳定的优秀表现.

Semi-supervised feature extraction method based on LPA-SKFST

In order to solve the problems of traditional LDA semi supervised feature extraction methods, such as the solution vector is not orthogonal, the solution space is unstable and there is no linear processing ability, this paper proposed LPA-SKFST. In this method, the LPA part increased the proportion of labeled samples through label propagation, semi supervised kernel optimal discriminant vectors SKFST combined KFST, Tikhonov regularization and global preserving regularization by two-way regularization method, and adopted the pair space solution method to ensure the uniform solution form when Fisher''s denominator matrix was singular or nonsingular. In the classification experiments of circle, iris, wine and pearl spectral, the accuracy of PCA, LDA, SLDA and SDG groups fluctuated with the change of sample set, labeled sample proportion and label reliability, while LPA-SKFST group kept stable above 85%. The results show that PA-SKFST can overcome the limitations of low proportion of labeled samples and unreliable labeling, and its performance is stable and excellent in both actual set and linear indivisible artificial set.