Abstract: The models of non-parametric kernel learning methods are generally built based on the manifold assumption and pairwise constraints. But it is so expensive for some intricate high-dimensional and sparse data, due to the high complexity of non-parametric kernel learning methods. In this paper, a sparse autoencoder with non-parametric method based on sparse self-coding is proposed. Through adding sparse autoencoder, the method not only overcomes overfitting problem and improves robustness, but also runs faster. The kernel cluster experiments are conducted with the kernel obtained by the proposed method. All results demonstrate that the proposed method outperforms the traditional non-parametric kernel learning method.

Key words: semi-supervised learning, pairwise constraints, autoencoder, sparse coding

摘要： 传统半监督非参核学习方法通常基于流形假设和成对约束信息建立学习模型。但是，这种模型对于某些复杂高维稀疏的数据而言算法复杂性较高。为了解决复杂高维稀疏数据核学习问题，提出一种基于稀疏自编码的非参核学习算法，通过稀疏自编码器引入稀疏约束，不仅提高了非参核学习方法的鲁棒性，避免了过拟合问题，而且提升了非参核学习算法的学习效率。通过核聚类实验验证了提出算法的有效性。实验结果表明，在非参核学习模型中融入了稀疏自编码器能够提高核聚类的效果，提升了半监督非参核学习算法的学习效率。

关键词: 半监督学习, 成对约束, 自编码器, 稀疏编码