Discriminative Model Selection for Belief Net Structures

  • Author(s) / Creator(s)
  • Technical report TR04-22. Model selection problem in Bayesian belief network (BN) structure learning is a classicial problem in the BN literature. To do model selection in BN structure learning, we need a evaluation score and a searching procedure. The generative criteria, AIC, BIC and BDe, are usually used to select the structure that is close to the true generative model. In this paper, we consider model selection in BN structure learning for classification tasks -- ie, we want to obtain the structure for an accurate BN classifier, As this is significantly different from the generative learning task, we consider using new discriminative criteria. Those discriminative criteria evaluate the classification performance of each structure instead of fitness to the joint distribution of the model. The discriminant criteria we investigated include Bias$^2$+Variance (BV), Classification Error (CE), and Conditional BIC (CBIC). Our experimental results suggest that, while discriminant model selection criteria generally performs better than discriminant model selection criteria, this is not universal. To understand why, we therefore investigated the model selection problem across different cases --- eg, with the true generative model having different complexity. Our experimental results suggest that this complexity of the true generative model does influence the performance of different model selection criteria. | TRID-ID TR04-22

  • Date created
  • Subjects / Keywords
  • Type of Item
  • DOI
  • License
    Attribution 3.0 International