The Evolution of Causal Models: A Comparison of Bayesian Metrics and Structure Priors

Created by W.Langdon from gp-bibliography.bib Revision:1.3973

@InProceedings{Neil:1999:ECM,
  author =       "Julian R. Neil and Kevin B. Korb",
  title =        "The Evolution of Causal Models: {A} Comparison of
                 {Bayesian} Metrics and Structure Priors",
  booktitle =    "Proceedings of the 3rd Pacific-Asia Conference on
                 Methodologies for Knowledge Discovery and Data Mining
                 ({PAKDD}-99)",
  year =         "1999",
  editor =       "Ning Zhong and Lizhu Zhou",
  volume =       "1574",
  series =       "Lecture Notes in Artificial Intelligence",
  pages =        "433--438",
  language =     "English",
  address =      "Beijing, China",
  month =        "26-28 " # apr,
  publisher =    "Springer-Verlag",
  keywords =     "genetic algorithms, genetic programming",
  ISBN =         "3-540-65866-1",
  DOI =          "doi:10.1007/3-540-48912-6_57",
  size =         "6 pages",
  abstract =     "We report the use genetic algorithms (GAs) as a search
                 mechanism for the discovery of linear causal models
                 when using two Bayesian metrics for linear causal
                 models, a Minimum Message Length (MML) metric [10] and
                 a full posterior analysis (BGe) [3]. We also consider
                 two structure priors over causal models, one giving all
                 variable orderings for models with the same arc density
                 equal prior probability (P1) and one assigning all
                 causal structures with the same arc density equal
                 priors (P2). Evaluated with Kullback-Leibler distance
                 prior P2 tended to produce models closer to the true
                 model than P1 for both metrics, with MML performing
                 slightly better than BGe. By contrast, when using an
                 evaluation metric that better reflects the nature of
                 the causal discovery task, namely a metric that
                 compares the results of predictive performance on the
                 effect nodes in the discovered model P1 outperformed P2
                 in general, with MML and BGe discovering models of
                 similar predictive performance at various sample sizes.
                 This supports our conjecture that the P1 prior is more
                 appropriate for causal discovery.",
  notes =        "p436 'For this study we employed a non-standard
                 genetic algorithm (cf. [5]), using DAGs directly as the
                 genetic representation, with genetic operators designed
                 for DAGs. These operators are described in detail in
                 [5] and are extended in [6].'

                 See also \cite{Neil96}",
}

Genetic Programming entries for Julian R Neil Kevin B Korb

Citations