大数据背景下贝叶斯模型平均的理论突破与应用前景

摘要

Model comparison and selection uncertainty issue is very common in the big data analysis. The Bayesian model averaging (BMA)treats model as stochastic variable and assigns prior and posterior probability for it in order to account for model uncertainty.BMA weights the results of each model by their posterior model probability,and in the end obtatin more robust results.In this paper,we briefly describe the origins and developments of BMA,introduce the paradigm of BMA,and then discuss new progresses of BMA.Some important aspects of application are given in the context of big data.BMA combined with complex data analysis methods will provide new insights in our big data research methods.%大数据统计分析过程中常面临模型比较和选择的不确定性问题。贝叶斯模型平均（BMA）方法可以通过先验和后验概率度量模型不确定性，并利用后验概率对模型的结果进行加权平均，最终得到更稳健的估计结果。在回顾贝叶斯模型平均发展历程的基础上，介绍贝叶斯模型平均的基本原理，综述其在一些难点问题上的理论进展，并介绍大数据背景下贝叶斯模型平均的应用前景。贝叶斯模型平均与复杂数据分析方法相结合，可能成为大数据研究的新思路。