一种基于逆模拟退火和高斯混合模型的半监督聚类算法

摘要

Semi-supervised Gaussian mixture model(SGMM)based on labeling nodes can improve the accuracy of model parameter estimation. However,the accuracy and convergence of the Expectation Maximization(EM)algorithm are affected by the amount of overlap and mixing coefficients among the Gaussian distributions. In order to improve the accu-racy and speed of the SGMM parameter estimation,the Anti-annealing is combined with the EM algorithm of SGMM. A clustering algorithm of the semi-supervised Gaussian mixture model based on anti-annealing(ASGMM-EM) is proposed. The inverse temperature parameter of the algorithm increases from a smaller value to an upper bound that more than 1 and then back to 1. The semi-supervised clustering EM algorithm is implemented at each inverse temperature parameter. Experiments on synthetic and real data show that the ASGMM-EM is better compared to the algorithms only using semi-supervised or anti-annealing technique.%基于节点标记的半监督高斯混合模型(Semi-supervised Gaussian Mixture Model,SGMM)可利用少量标记样本提高模型参数估计的准确率,但参数估计算法(SGMM Expectation Maximization,SGMM-EM)的准确率和收敛速度受高斯分布之间的重叠度和混和系数差异度影响.为提高SGMM模型参数估计的准确率和收敛速度,将逆模拟退火框架与SGMM模型的EM算法相结合,提出一种基于逆模拟退火框架的半监督高斯混合模型聚类算法(Anti-annealing SGMM-EM,ASGMM-EM).该算法逆温度参数从一个较小且大于0的值逐渐增加到大于1的上界,再逐渐降回1.在每个逆温度参数下执行半监督聚类算法SGMM-EM并迭代至收敛.人工数据和真实数据上实验表明提出的算法ASGMM-EM优于仅用半监督技术或逆模拟退火技术的基于高斯混合模型的EM算法.