Subspace Classification Based On Bayesian Hierarchical Model

This paper discusses a subspace classification method based on bayesian hierarchical model. This particular method possesses solid foundation on statistics comparing to traditional subspace classification methods. Each attribute is given certain probability to decide whether the attribute is generated by unite model or single model. Therefore, in our paper the subspace is constructed by the different weighted attributes.Based on Bayesian Hierarchical Model, the model-based approach allow us to realize and evaluate model in cognitive science. Different parameters are separated on different levels in order to divide a complex estimated problem into some easily estimated problems. It can describe the model in detail,having good generalization and applicability.Our work involved in model behavior design, inference, classification de-cision,parameter estimation based on MCMC algorithms and introduces the "mean shift" principle to reduce high dimension to lower dimension for analysis. UCI datasets are used to testify the classfication accuracy of the model. Furthermore, we apply the model to spatially correlated genetic array-CGH data in which we try to identify groups of tumor cells having common patterns of chromosomal abnormalities. Via the analysis of array data, we summarize advantage of the model when classify this data:1) the model is generating model having good congitive process when modelling;2) excellent classification accuracy when classify array-CGH data;3) this model can identify which genes are differentially expressed across different kinds of tissue samples.Finally, summarizing some innovation in our paper, proposing variational method and dirichlet mixture model to further modify our model.