鲁棒贝叶斯混合分布模型选择.pdfVIP

( ) 45 5 Vol. 45, No. 5 JOURNAL OF NANJING U NIVERSITY 2009 9 Sept. 2009 ( NAT URA L SCIENCES) * ** 卿湘运, 王行愚 ( , , 2002 7) : ( deriance informat ion criterion, DIC) . , DIC ; , . , Old Faithful Geyser : . : , , , , : T P 9 Model selection for robust Bayesian mixture distributions Qing X iangYun, Wang X ingYu ( School of Information Science and Engineering, East China University of Science and T echnology, Shanghai, 2002 7, China) Abstract: Bayesian approaches to robust mixture modelling based on Studentt distributions enable to be less sensitive to outliers, thereby prevent ing from overestimat ing of the number of mixting component s. H ow ever, there are tw o intractable problems in the previous methods for model selection under the variat ional Bayesian framew ork: ( 1) The variational approach converges to a local maximum of the low bound on t he logevidence that dependents on the init ial parameter values. How can the variat ional approach guarantee that t he initial settings for different models are consistency ? (2) T he low bound is sensitive t o factorized approximation forms in the inference process. H ow can the variat ional approach guarant ee that the approximate errors for different models are equivalent? In this paper, we present a model selection algorit hm for robust bayesian mixture distributions based on devian