- author = "Richard J. Gilbert and Royston Goodacre and Beverly Shann and Douglas B. Kell and Janet Taylor and Jem J. Rowland",
- title = "Genetic Programming-Based Variable Selection for High-Dimensional Data",
- booktitle = "Genetic Programming 1998: Proceedings of the Third Annual Conference",
- year = "1998",
- editor = "John R. Koza and Wolfgang Banzhaf and Kumar Chellapilla and Kalyanmoy Deb and Marco Dorigo and David B. Fogel and Max H. Garzon and David E. Goldberg and Hitoshi Iba and Rick Riolo",
- pages = "109--115",
- address = "University of Wisconsin, Madison, Wisconsin, USA",
- publisher_address = "San Francisco, CA, USA",
- month = "22-25 " # jul,
- publisher = "Morgan Kaufmann",
- keywords = "genetic algorithms, genetic programming",
- ISBN = "1-55860-548-7",
-
URL = "
http://dbkgroup.org/Papers/Gilbert%20et%20al%201998%20Genetic%20programming-based%20variable%20selection.pdf",
-
URL = "http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/gp1998/gilbert_1998_GPvshdd.pdf",
- size = "7 pages",
-
abstract = "A major advantage of the genetic programming [GP]
approach to data modeling is the automatic ability of
the GP to select input variables that contribute
beneficially to the model and to disregard those that
do not. GPs are thus able to reduce substantially the
dimensionality of the model, with consequent
interpretation benefits. Experimental analytical
techniques frequently generate data with very high
dimensionality, typically measuring many tens or even
hundreds of variables per sample. It is often not
apparent which of the measured variables can best be
used to derive a predictive model describing the data.
The identification of these variables often provides a
better understanding of the physical, chemical or
biological mechanism underlying the experimental
observations.
The ability of a GP to perform variable selection is assessed with regard to a binary classification of the sporulation state of bacterial strains. The analytical technique used, Curie-point pyrolysis mass spectrometry, generates data for 150 variables per sample. The GP-derived predictive rules for the~e data contain a substantially smaller subset of these variables, typically just 6-9. Inspection of these rules leads to the somewhat counter-intuitive conclusion that the best predictive models use both highly characteristic and highly non-characteristic variables.",
- notes = "GP-98",
