calcDependence: Dependence with target variable
In ddgraph: Distinguish direct and indirect interactions with Graphical Modelling

Description Usage Arguments Details Value References Examples

Calculate dependence with a target variable

1	calcDependence(dd, method="ncpc", ...)

dd

An object of type DDDataSet

method

Algorithm to use. Valid values are:

ncpc - Neighbourhood Consistent PC algorithm
ncpc* - Neighbourhood Consistent PC algorithm star version
hc - Hill-climbing with custom penalty functions
hc-bic - Hill-climbing with BIC penalization (package bnlearn)
hc-bde - Hill-climbing with BDe penalization (package bnlearn)
iamb - IAMB algorithm (package bnlearn)
fast.iamb - FastIAMB algorithm (package bnlearn)
inter.iamb - InterIAMB algorithm (package bnlearn)
pc - PC algorithm (package pcalg)
mmpc - MMPC algorithm (package bnlearn)
mmhc - MMHC with custom penalty functions
mmhc-bic - MMHC with BIC penalization (package bnlearn)
mmhc-bde - MMHC with BDe penalization (package bnlearn)

...

Extra parameters passed to backend functions ncpc(), plotBNLearn() and plotPCAlgo() depending on the picked algorithm (parameter method).

Extra parameters for ncpc and ncpc*:

alpha - the alpha (P-value) cutoff for conditional independence tests (default: 0.05)
p.value.adjust.method - the multiple testing correction adjustment method (default: "none")
test.type - the type of conditional independence test (default: "mc-x2-c"). See the documentation for ciTest for available conditional independence tests
max.set.size - the maximal number of variables to condition on, if NULL estimated from number of positives in class labels. Needs to be specified for continuous data. (default: NULL)
mc.replicates - the number of Monte-Carlo replicates for the conditional independence test, if applicable (default: 5000)
report.file - name of the file where a detailed report is to be printed, reporting is suppressed if NULL (default: NULL)
verbose - if to print out information about how the algorithm is progressing (default: TRUE)
min.table.size - the minimal number of samples in a contingency table per conditioning set (applicable only for discrete data) (default: 10)

Extra parameters for hc, mmhc:

score - score function to use, accepts all from bnlearn package. For discrete data: "loglik", "aic", "bic", "bde", "k2". For continuous: "loglik-g", "aic-g", "bic-g", "bge". For more details see help page for package-bnlearn.
make.plot - if to make a plot or just return the network (default: FALSE)
blacklist - a data frame with two columns (optionally labeled "from" and "to"), containing a set of arcs not to be included in the graph. (default: NULL)
restart - the number of random restarts for score-based algorithms (default: 0)
scale - the colour scaling (default: 1.5)
class.label - the label to use for the target variable (default: "target")
use.colors - if to colour code the enrichment/depletion in a plot (default: TRUE)

Extra parameters for hc-bic, hc-bde, mmhc-bic, mmhc-bde:

make.plot - if to make a plot or just return the network (default: FALSE)
blacklist - a data frame with two columns (optionally labeled "from" and "to"), containing a set of arcs not to be included in the graph. (default: NULL)
restart - the number of random restarts for score-based algorithms (default: 0)
scale - the colour scaling (default: 1.5)
class.label - the label to use for the target variable (default: "target")
use.colors - if to colour code the enrichment/depletion in a plot (default: TRUE)

Extra parameters for iamb, fast.iamb, inter.iamb, mmpc:

make.plot - if to make a plot or just return the network (default: FALSE)
alpha - the alpha value of conditional independence tests (default: 0.05)
test - the type of conditional independence test (default: "mc-mi"). For conditional independence tests available consult the bnlearn package help page (?bnlearn).
B - the number of Monte-Carlo runs for conditional independence tests, if applicable (default: 5000)
blacklist - a data frame with two columns (optionally labeled "from" and "to"), containing a set of arcs not to be included in the graph. (default: NULL)
scale - the colour scaling (default: 1.5)
class.label - the label to use for the target variable (default: "target")
use.colors - if to colour code the enrichment/depletion in a plot (default: TRUE)

Extra parameters for pc:

alpha - the alpha value cut-off for the conditional independence tests (default: 0.05)
verbose - if to show progress (default: FALSE)
directed - if TRUE applies PC algorithm, if FALSE applies PC-skeleton (default: TRUE)
make.plot - if to make a plot of the final inferred network (default: FALSE)
scale - the scaling parameter for color-coding (default: 1.5)
indepTest - the independence test wrapper function (default: mcX2Test). The following functions are available: mcX2Test (a wrapper around mc-x2-c (Monte Carlo X2 test) with B=5000), mcX2TestB50k (a wrapper around mc-x2-c (Monte Carlo X2 test) test with B=50000), mcMITest (wrapper around mc-mi test from bnlearn with B=5000). The package pcalg additionally provide following tests: binCItest for binary data (performs a G^2 test) and gaussCItest for continuous data (performs Fisher's Z transformation), dicCItest for discrete data (performs G^2 test).
class.label - the label to show for target variable (default: "target")
use.colors - if to colour code the results (default: TRUE)

This function is a front-end convenience function to access predictions of direct dependence with a target variable by various Graphical Modelling algorithm.

Consider a set of variable X_1, ..., X_m and a target variable T. We say that that X_i is directly dependent with T if there is no other set of variable X_j, X_k, ... such that it renders X_i conditionally independent of T. In other words, X_i is the most immediate casual cause/consequence of T in the set of chosen variables.

Note that the above statement is different from that of classical feature selection for classification. A set of features obtained with feature selection have the property that a good classifier can be made based on them alone, while the above statement establishes statistical properties of variables. The set of variables with direct dependence might not be optimal for classification, since classification performance can be strongly influenced by false negatives (Friedman et al, 1997).

A list with elements:

obj - the resulting object, either of class DDGraph for ncpc and ncpc* algorithms, or of class bn for bnlearn algorithms, or of class pcAlgo for PC algorithm.
nbr - the variables with direct dependence (i.e. target node neighbourhood in the causal graph). For both ncpc and ncpc* includes variables with direct and joint dependence.
mb - the variables in Markov Blanket of target variable. Not applicable for ncpc algoritm. For ncpc* algorithm includes variables with direct, joint and conditional dependence.
labels - for ncpc and ncpc* contains the set of labels that are output of the algorithm.

Nir Friedman, Dan Geiger, and Moises Goldszmidt, "Bayesian Network Classifiers", Machine Learning 29 (November 1997): 131-163.

# load in the data for fly mesoderm
data(mesoBin)

# increase alpha to 0.1, suppress progress output
calcDependence(mesoBin$VM, "ncpc", alpha=0.05)

# run ncpc* with mutual information with shrinkage and minimal numbers of 
# samples per conditioning set of 15
calcDependence(mesoBin$VM, "ncpc*", test.type="mi-sh", min.table.size=15)

# run PC algorithm using the G^2 test from pcalg package
calcDependence(mesoBin$VM, "pc", indepTest=pcalg::binCItest)

# run hill-climbing with BIC penalization and plot the resulting Bayesian Network
# NOTE: plotting requires the Rgraphviz package
if(require("Rgraphviz"))
calcDependence(mesoBin$VM, "hc-bic", make.plot=TRUE)

# continuous data example
data(mesoCont)

# run ncpc with linear correlation test and with maximal conditioning set of 3
res <- calcDependence(mesoCont$VM, "ncpc", max.set.size=3, test.type="cor")
# plot the resulting ddgraph with colours
if(require("Rgraphviz"))
plot(res$obj, col=TRUE)