Maximum likelihood estimation and the EM algorithm:
Detailed EM algorithm and Gaussian mixture model (Gaussian mixture model, GMM) _ Lin Limin love taking a bath -CSDN blog https://blog.csdn.net/lin_limin/article/details/81048411
GMM (Gaussian mixture model) and a simple realization column -CSDN blog _zjm750617105 of https://blog.csdn.net/zjm750617105/article/details/5243029
A fraction choice
In a GMM, BIC (Bayesian Information Criteria) criterion is an effective method to determine the score. Note that if you use a Variational Bayesian Gaussian mixture
avoid designated as a fraction of the GMM.
AIC criterion
AIC guidelines by the Japanese in 1973 and statisticians Akaike proposed, the full name is to minimize the amount of information criterion (Akaike Information Criterion). It is fitting accuracy and number of parameters weighting function: AIC=2(模型参数的个数)-2ln(模型的极大似然函数)
BIC criterion
AIC provides effective rules for model selection, but there are also disadvantages. When the sample size is large, AIC criterion fitting error information will be provided in an enlarged sample size, the number of parameters and the penalty factor, but the sample size does not matter (always 2), so when a large sample size , using the AIC criteria for selecting model does not converge with the real model, which is usually more than the number of unknown parameters contained in the real model. BIC (Bayesian InformationCriterion) Bayesian Information Criterion is a criterion proposed by Schwartz in 1978, according to Bayes theory, known as SBC guidelines (also known as BIC), up for the lack of AIC. SBC is defined as: BIC = ln(n)(模型中参数的个数) - 2ln(模型的极大似然函数值)
【sklearn第十九讲】高斯混合模型_python_wong2016的博客-CSDN博客 https://blog.csdn.net/wong2016/article/details/81057023
二维高斯分布(Two-dimensional Gaussian distribution)的参数分析_林立民爱洗澡-CSDN博客 https://blog.csdn.net/lin_limin/article/details/81024228