[UAI] Question: Overfitting in BN Parameter Learning

From: Han-Shen Huang (julio@hugo.csie.ntu.edu.tw)
Date: Wed Feb 27 2002 - 09:05:26 PST

  • Next message: Colin Campbell, Engineering Mathematics: "[UAI] Special Issue/Analysis of Microarray Data"

    Dear All,

    Suppose we have an incomplete data set D and a known Bayesian
    network structure Bs, and use EM to learn the parameter for
    Bs from D. It is possible that the final parameter has longer
    KL distance to the answer than other parameters obtained during
    the EM process.

    I would like to seek your helps for the following questions:
    1. Is there any related work in the topic?
    2. Can we call the situation "parameter overfitting"? why not
       "underfitting" or other terms? If I want to use minimum description
       length principle as the criterion to choose the parameter, does the
       description length of the BN differs when I use different parameters
       but fix the structure?

    Thank you very much.

    ------------------------------------------------
    Han-Shen Huang
    Ph.D. Student,
    Department of CSIE, National Taiwan University
    E-mail: hshuang@hugo.csie.ntu.edu.tw

    This archive was generated by hypermail 2b29 : Wed Feb 27 2002 - 09:09:59 PST