R help - Apr 2013 - bnlearn: how to compute boot strength with mmhc and a blacklist

Dear R-help list:

I have two related questions regarding the functions boot.strength and 
custom.strength in bnlearn.

1)
I am using the following commands (on a set of continuous data, the 
example here run on fake data):

 > myblacklist<-data.frame(from=c("x1", "x1", 
"x1",   "x2", "x2",
"x2")) ,  to=c("x2", "n3",  "n4", 
"x1"," n3", "n4"))
 > result <- boot.strength(data, R=200, algorithm="mmhc", 
algorithm.args=list(blacklist=myblacklist))

 > bootstrength
     from    to    strength direction
1    n3    n4    1.000       0.5
2    n3    x1    0.340       0.5
3    n3    x2    0.115       0.5
4    n4    n1    1.000       0.5
5    n4    x1    0.080       0.5
6    n4    x2    1.000       0.5
7    x1    n3    0.340       0.5
8    x1    n4    0.080       0.5
9    x1    x2    0.000       0.0
10    x2    n3    0.115       0.5
11    x2    n4    1.000       0.5
12    x2    x1    0.000       0.0


Question: I have specified a blacklist. I would have expected this to 
completely disallow the arcs on the blacklist. But the result shows that 
some of the blacklisted arcs have a strength > 0 (rows 7,8,10,11). It 
seems that only the arc that was blacklisted in both directions was 
actually banned (x1-x2, in rows 9 and 12). What is the reason for this? 
Is there a way to completely disallow all blacklisted arcs, such that 
their strength is 0.0? Or is there a compelling reason why that should 
not be done?


2) In the documentation of custom.strength, the following code example 
is given:

start = random.graph(nodes = names(learning.test), num = 50)
netlist = lapply(start, function(net) { hc(learning.test, score =
"bde",
iss = 10, start = net) })
arcs = custom.strength(netlist, nodes = names(learning.test), cpdag = FALSE)

This code makes 50 different networks from the same data, then uses them 
as input for custom.strength. The networks are constructed using the 
algorithm "hc". A different network is produced every time
"hc" is
invoked because a random starting network is supplied to the parameter 
"start". I would like to do the same thing, but use "mmhc"
instead of
"hc". However, in my hands, the networks that are constructed by
"mmhc"
are all identical, and I am not sure how to introduce a random element 
into the construction. Question: Which parameters do I need to give to 
"mmhc" in order to obtain a different network every time it is run on 
the same data set?

Any help is greatly appreciated!

-- 
Leonore Wigger
University of Lausanne



	[[alternative HTML version deleted]]

Dear Leonore,

On Wed, Apr 10, 2013 at 5:51 PM, Leonore Wigger <leonore.wigger at
unil.ch> wrote:
> Question: I have specified a blacklist. I would have expected this to
> completely disallow the arcs on the blacklist. But the result shows that
> some of the blacklisted arcs have a strength > 0 (rows 7,8,10,11). It
> seems that only the arc that was blacklisted in both directions was
> actually banned (x1-x2, in rows 9 and 12). What is the reason for this?
> Is there a way to completely disallow all blacklisted arcs, such that
> their strength is 0.0? Or is there a compelling reason why that should
> not be done?
Because by default boot.strength() runs with "cpdag = TRUE". This
means that reversible arcs can have positive strength in both
directions. You should set "cpdag = FALSE" to get the result you are
expecting. In that case the probabilities of the arc directions should
be taken with a grain of salt, as they can be influenced by many
things (optimized = TRUE/FALSE, order of the variables in the data
set) unless you are doing causal modelling.
> This code makes 50 different networks from the same data, then uses them
> as input for custom.strength. The networks are constructed using the
> algorithm "hc". A different network is produced every time
"hc" is
> invoked because a random starting network is supplied to the parameter
> "start". I would like to do the same thing, but use
"mmhc" instead of
> "hc". However, in my hands, the networks that are constructed by
"mmhc"
> are all identical, and I am not sure how to introduce a random element
> into the construction. Question: Which parameters do I need to give to
> "mmhc" in order to obtain a different network every time it is
run on
> the same data set?
This is not surprising, because mmhc() does not have a "start"
argument, so it's starting from the same network over and over. There
is no way to provide a random seed to mmhc(), so the only way to
perturb it is through bootstrap.

Marco

-- 
Marco Scutari, Ph.D.
Research Associate, Genetics Institute (UGI)
University College London (UCL), United Kingdom