abn (version 2.2.2)

mostprobable: Find most probable DAG structure

Description

Find most probable DAG structure using exact order based approach due to Koivisto and Sood, 2004

Usage

mostprobable(score.cache, prior.choice=1, score="mlik", verbose=TRUE)

Arguments

score.cache

object of class abnCache typically outputted by from buildscorecache().

prior.choice

an integer, 1 or 2, where 1 is a uniform structural prior and 2 uses a weighted prior, see details

score

character giving which network score should be used to select the structure.

verbose

if TRUE then provides some additional output.

Value

An object of class abnMostprobable, which is a list containing: a matrix giving the DAG definition of the most probable posterior structure, the cache of pre-computed scores and the score used for selection.

Details

The procedure runs the exact order based structure discovery approach of Koivisto and Sood (2004) to find the most probable posterior network (DAG). The local.score is the node cache, as created using buildscorecache (or manually provided the same format is used). Note that the scope of this search is given by the options used in local.score, for example, by restricting the number of parents or the ban or retain constraints given there.

This routine can take a long time to complete and is highly sensitive to the number of nodes in the network. It is recommended to use this on a reduced data set to get an idea as to the computational practicality of this approach. In particular, memory usage can quickly increase to beyond what may be available. For additive models, problems comprising up to 20 nodes are feasible on most machines. Memory requirements can increase considerably after this, but then so does the run time making this less practical. It is recommended that some form of over-modeling adjustment is performed on this resulting DAG (unless dealing with vast numbers of observations), for example, using parametric bootstrapping, which is straightforward to implement in MCMC engines such as JAGS or WinBUGS. See the case studies at http://r-bayesian-networks.org or the files provided in the package directories inst/bootstrapping_example and inst/old_vignette for details.

The parameter prior.choice determines the prior used within each node for a given choice of parent combination. In Koivisto and Sood (2004) p.554, a form of prior is used, which assumes that the prior probability for parent combinations comprising of the same number of parents are all equal. Specifically, that the prior probability for parent set G with cardinality |G| is proportional to 1/[n-1 choose |G|] where there are n total nodes. Note that this favors parent combinations with either very low or very high cardinality, which may not be appropriate. This prior is used when prior.choice=2. When prior.choice=1 an uninformative prior is used where parent combinations of all cardinalities are equally likely. The latter is equivalent to the structural prior used in the heuristic searches e.g., searchHillclimber or searchHeuristic.

Note that the network score (log marginal likelihood) of the most probable DAG is not returned as it can easily be computed using fitabn, see examples below.

References

Koivisto, M. V. (2004). Exact Structure Discovery in Bayesian Networks, Journal of Machine Learning Research, vol 5, 549-573.

Further information about abn can be found at: http://r-bayesian-networks.org

Examples

# NOT RUN {
##############################
## Example 1
##############################

## this data comes with abn see ?ex1.dag.data
mydat <- ex1.dag.data 

## setup distribution list for each node
mydists <- list(b1="binomial", p1="poisson", g1="gaussian", b2="binomial",
              p2="poisson", b3="binomial", g2="gaussian", b4="binomial",
              b5="binomial", g3="gaussian")

## assum no constraints in ban nor retain

## parent limits
max.par <- list("b1"=4,"p1"=4,"g1"=4,"b2"=4,"p2"=4,"b3"=4,"g2"=4,"b4"=4,"b5"=4,"g3"=4)
## now build cache
mycache <- buildscorecache(data.df = mydat, data.dists = mydists, max.parents = max.par)

## Find the globally best DAG
mp.dag <- mostprobable(score.cache=mycache)
fitabn(object=mp.dag)$mlik

## plot the best model
myres <- fitabn(object=mp.dag,create.graph=TRUE)

plotabn(dag.m = mp.dag$dag, data.dists = mydists)#, fitted.values.abn = myres)

## fit the known true DAG
true.dag <- matrix(data=0,ncol=10, nrow = 10)
colnames(true.dag) <- rownames(true.dag) <- names(mydists)

true.dag["p2",c("b1","p1")] <- 1
true.dag["b3",c("b1","g1","b2")] <- 1
true.dag["g2",c("p1","g1","b2")] <- 1
true.dag["b4",c("g1","p2")] <- 1
true.dag["b5",c("g1","g2")] <- 1
true.dag["g3",c("g1","b2")] <- 1

fitabn(dag.m = true.dag, data.df = mydat, data.dists=mydists)$mlik

#################################################################
## Example 2 - models with random effects
#################################################################

## this data comes with abn see ?ex3.dag.data
# mydat <- ex3.dag.data[,c(1:4,14)] 

# mydists <- list(b1="binomial", b2="binomial", b3="binomial",
#              b4="binomial")

## This takes a few seconds
# mycache.mixed <- buildscorecache(data.df = mydat, data.dists = mydists,
#               group.var = "group", cor.vars = c("b1","b2","b3","b4"),
#               max.parents=2, which.nodes=c(1:4))

## find the most probable DAG
# mp.dag <- mostprobable(score.cache=mycache.mixed)

## get goodness of fit
# fitabn(object = mp.dag, data.df = mydat, data.dists = mydists,
#      group.var = "group", cor.vars = c("b1","b2","b3","b4"))$mlik
# }