Identifying character non-independence in phylogenetic data using data mining techniques

摘要

Undiscovered relationships in a data set may confound analyses, particularly those that assume data independence. Such problems occur when characters used for phylogenetic analyses are not independent of one another. A main assumption of phylogenetic inference methods such as maximum likelihood and parsimony is that each character serves as an independent hypothesis of evolution. When this assumption is violated, the resulting phylogeny may not reflect true evolutionary history. Therefore, it is imperative that character non-independence be identified prior to phylogenetic analyses. To identify dependencies between phylogenetic characters, we applied three data mining techniques: 1) Bayesian networks, 2) decision tree induction, and 3) rule induction from coverings. We briefly discuss the main ideas behind each strategy, show how each technique performs on a small sample data set, and apply each method to an existing phylogenetic data set. We discuss the interestingness of the results of each method, and show that, although each method has its own strengths and weaknesses, rule induction from coverings presents the most useful solution for determining dependencies among phylogenetic data at this time.