Efficient estimation exploiting independence constraints
- Russ Greiner, Dept of Computing Science; PI of AICML
- Dale Schuurmans, AICML
- Chris O'Brien
There is an obvious way to obtain an unbiased estimate of the response to a query P ( C = c j E= e ) from a sufficiently large data sample: just use the empirical frequency. Unfortunately, this can have high variance (or even be undefined) when the conditioning event E = e is rare. In some situations, we are able to reduce this variance by exploiting our knowledge about the underlying distribution --- in particular, by using a known set of probabilistic indepedencies amongst the variables, encoded in the structure of a Bayesian belief network. The report investigates ways to exploit those independencies when estimating the response to a probabilistic query. In particular, we show a general approach that produces an unbiased estimate with small variance. Of independent interest, we also present a closed form expression for the (asymptotic) variance of a response wrt a given structure and sample.
Citation
R. Greiner, D. Schuurmans, C. O'Brien. "Efficient estimation exploiting independence constraints". January 2002.Keywords: | machine learning |
Category: |
BibTeX
@incollection{Greiner+al:02, author = {Russ Greiner and Dale Schuurmans and Chris O'Brien}, title = {Efficient estimation exploiting independence constraints}, year = 2002, }Last Updated: March 07, 2007
Submitted by Nelson Loyola