基于核自调整进行半监督聚类

Kernel-based adaptation for semi-supervised clustering

半监督聚类是通过在无监督算法的基础上加入有限的背景知识来实现的。现有的基于核的半监督聚类算法对于核参数的设定仍需人工进行调节,其选择值会极大地影响最终的结果。通过将关联加入到聚类目标函数中,在聚类过程反复地优化高斯核参数,自动确定最佳RBF核,并将最佳核计算与SSKK算法结合起来得到SSKKOK算法。实验结果表明,该算法能在利用基于核半监督聚类算法功能的基础上自动设置有关的参数。

Semi-supervised clustering combines unsupervised clustering algorithms with limited background knowledge. However, setting of kernel parameters involved in SSKK（ semi-supervised kernel-based kmeans） is still needed to manually regularize. The chosen value of critical parameters will largely effect on the result. The pair constraints were incorporated into clustering object function, and kernel parameters were optimized iteratively during clustering process till to determine optimal RBF kernel. Furthermore, SSKKOK algorithm was proposed, which was based on optimal kernel computation and SSKK algorithm. The experiment demonstrates the algorithm can not only utilize the effect of SSKK , but also automatically set involved parameters.