Parameter learning for BN

[kenneth-lee-ch]

I think the functions computeCPTfromDF found on (https://pyagrum.readthedocs.io/en/1.9.0/notebooks/17-Examples_parametersLearningWithPandas.html) will throw an error based on the following examples

def computeCPTfromDF(bn,df,name):
    """
    Compute the CPT of variable "name" in the BN bn from the database df
    """
    id=bn.idFromName(name)
    parents=list(reversed(bn.cpt(id).names))
    domains=[bn.variableFromName(name).domainSize()
             for name in parents]
    parents.pop()

    if (len(parents)>0):
        c=pd.crosstab(df[name],[df[parent] for parent in parents])
        s=c/c.sum().apply(np.float32)
    else:
        s=df[name].value_counts(normalize=True)

    bn.cpt(id)[:]=np.array((s).transpose()).reshape(*domains)

def ParametersLearning(bn,df):
    """
    Compute the CPTs of every varaible in the BN bn from the database df
    """
    for name in bn.names():
        computeCPTfromDF(bn,df,name)

Examples

bn=gum.fastBN("A[10]->B[10]<-C[10]")
print(bn.cpt("A")* bn.cpt("B") * bn.cpt("C"))
gum.generateSample(bn, 100, name_out ="random.csv")
df = pd.read_csv("random.csv")
ParametersLearning(bn,df)
print(bn.cpt("A")* bn.cpt("B") * bn.cpt("C"))

[phwuil]

You are right but I am not familiar enough with pandas to correct it :-)

The problem is that you generate only 100 samples for a joint of size 10x10x10 => obvisouly, you do not have all the cases for the joint. Then the crosstab does not contain all the cases and has a size < 1000. Then you cannot reshape it.

For instance, with only 10 generated samples, here is the crosstab :

As you can see, B=1 is never seen, (A,C) in {(0,0),(0,1),(1,0),...} neither ... So the crosstab have a size 70 that we want to reshape in 10x10x10. Leading to the error :

If you generate enough sample (say 10 000), the generation explore all the possible events and then the crosstab contains 10x10x10 elements and every thing is OK :-)

The solution should be to add columns and rows of 0.0 for all the missing event... But I do not know how to do that in pandas

[phwuil]

Note that it is much more convenient to use pyAgrum.BNLearner.fitParameters :-)

But you stil have the problem with a too small sample size

Except if you use use a prior or a smoothing as proposed in the error above :

[kenneth-lee-ch]

What is the difference between the functions I posted and the pyAgrum.BNLearner.fitParameters?

…

On Thu, Aug 17, 2023, 6:22 AM Pierre-Henri Wuillemin < ***@***.***> wrote: Note that it is much more convenient to use pyAgrum.BNLearner.fitParameters :-) [image: image] <https://user-images.githubusercontent.com/5980636/261279002-223c3c88-dce2-4cb9-9c8a-f458e126eb18.png> But you stil have the problem with a too small sample size [image: image] <https://user-images.githubusercontent.com/5980636/261279159-3cd28d2a-d196-43f9-ab21-5549eba3f581.png> Except if you use use a prior or a smoothing as proposed in the error above : [image: image] <https://user-images.githubusercontent.com/5980636/261279531-36976856-706c-4d54-b50d-ff7f03a20433.png> — Reply to this email directly, view it on GitHub <#19 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF3Z3CS5KPBBIHXPESCDMSDXVXWFDANCNFSM6AAAAAA3TTQZCY> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[phwuil]

fitParameters does not use panda. This notebook is just an example to understand how parameter learning works but it is not a correct way to do it.

fitParameters (or the less convenient learnParameters) can include prior or smoothing, is multithreaded and written in C++ to be as fast as possible.

[kenneth-lee-ch]

So, if I don’t use prior smoothing, the resulting potential will be the same?

…

On Aug 17, 2023, at 9:43 AM, Pierre-Henri Wuillemin ***@***.***> wrote: fitParameters does not use panda. This notebook is just an example to understand how parameter learning works but it is not a correct way to do it. fitParameters (or the less convenient learnParameters) can include prior or smoothing, is multithreaded and written in C++ to be as fast as possible. — Reply to this email directly, view it on GitHub <#19 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF3Z3CUFRCCJXB2A4XIXAQTXVYNZBANCNFSM6AAAAAA3TTQZCY>. You are receiving this because you authored the thread.

[phwuil]

It should :-)

But, once again , for your example of 100 samples, there is no way to fit the parameters correctly without priors or smoothing

[kenneth-lee-ch]

But I guess using fit.parameter will automatically fill in 0 for the missing events?

…

On Aug 17, 2023, at 10:02 AM, Pierre-Henri Wuillemin ***@***.***> wrote: It should :-) — Reply to this email directly, view it on GitHub <#19 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF3Z3CTMARZTE6JPHHF5UXTXVYP7JANCNFSM6AAAAAA3TTQZCY>. You are receiving this because you authored the thread.

[phwuil]

The problem is when a conditionning set is not observed.
Let say that A[3] is never observed with the value 1.

P(A=1)=0 can be correctly computed from the samples. (but is this correct ? proba=0 means never )

but if A->B in the BN, how can we computed P(B|A=1) since we never observed values of B when A is 1 ...
P(B|A=1)=[0,0] ...

• With smoothing, P(B|A=1)=[0.5 , 0.5 ]
• With prior (Dirichlet for instance) you can specify more precisely this distribution