中文摘要：

核函数及其参数的选择决定着核方法的性能．本文基于半监督学习思想，通过构建一个目标函数，利用无标签数据和成对约束信息来优化核函数，使得核函数尽可能适应数据集，从而改善核函数性能．为验证方法的有效性，将其应用于核主成分分析（KPCA）的核函数优化中，在人工数据和UCI数据集上对KPCA提取特征的分类和聚类性能进行评估，实验结果说明提出方法改进了分类和聚类性能．

英文摘要：

The selection of kernel function and its parameters determine the performance of kernel function. A semi-supervised data-dependent kernel optimization algorithm is presented, which uses unlabeled data and pairwise constraints to maximize an objective function sensitive to data-dependent kernel, so that its performance is improved. Then the proposed method is employed to optimize the kernel of kernel principal components analysis （KPCA） and the experimental results of the classification and clustering performance on the artificial data and UCI data sets show its efficiency.