max-min-hill-climbing-algorithm

The max-min hill-climbing Bayesian network structure learning algorithm, Ioannis Tsamardinos · Laura E. Brown · Constantin F. Aliferis, Mach Learn DOI 10.1007/s10994-006-6889-7

*This algorithm reconstructs Bayesian Networks from observational data. Therefore it first builds the skeleton of the DAG (directed acyclic graph) with the max-min parents and children (MMPC) algorithm. Afterwards it directs the edges between the vertices with the Bayesian Dirichlet likelihood-equivalence uniform score.

For more information on that read the report appended or* The max-min hill-climbing Bayesian network structure learning algorithm , by Ioannis Tsamardinos, Laura E. Brown & Constantin F. Aliferis.

INSTALLATION

Before you can use this package, be sure you have the latest R version (>=3.1), RCPP version (>=0.11.1) and the igraph package installed.

Download the R source file (mmhc_1.0.tar.gz), open R (in the console, RStudio, etc.) and install the package into your R environment with:

install.packages("mmhc_1.0.tar.gz")

Include the library with:

library(mmhc)

Here is the example from the man pages of the package:

Basic R functions

• student(int x) returns a data frame with x observations; one of the two self made examples
• rainy(int x) returns a data frame with x observations; one of the two self made examples
• mmhc(data.frame x) input: the observed data x (a data frame); calculates everything and makes a plot

Basic C++ methods

• C <- new(MMHC, data.frame) initalizes the mmhc class
• C$mmpc() executes the MMPC algorithm
• C$mmhc() executes the MMHC algorithm
• C$pc() this C++-methods returns the PC set
• C$adjMat() this C++-methods returns the adjacency matrix
• C$score() this C++-methods returns the score of the reconstructed graph
• C$mat() this C++-methods returns the data frame converted into an integer matrix

How to use the algorithm

• data <- student(1000) # as above
• mmhc(data) # gives you the plot of the graph (no return value)

Manuel Workflow

Producing the data step by step

• library(Rcpp) # load Rcpp package
• library(igraph) # load igraph package
• data <- student(1000) # initalize the underlying example with 1000 observations
• C <- new(MMHC, data) # initalize the class object
• C$mmpc() # first reconstruct the skeleton (max-min parents and children algorithm)
• C$pc() # returns the PC set. It is a list where the n-th list element stands for the n-th node in your graph. The elements of one node are the parents/children of the node.
• C$mmhc() # set the edges (BDeu score)
• C$adjMat() # returns the adjacency matrix
• C$score() # returns the score of the graph
• plotObj <- graph.adjacency(C$adjMat()) # makes a plotable object with the igraph package
• plot(plotObj) # plots the object