How can I manipulate the way conditional probability table is formatted?

[kenneth-lee-ch]

Suppose I have some random BN generated and I generate some conditional probability table using cpt(). Given a string, how can I always be able to get a table where the given string is on the column instead of rows?

For example, here I attached an example, let's say I want to have B to be where D is always for any given bn.cpt() whenever B exists in the cpt.

Is there a way to do that for any given cpt?

BEFORE:

AFTER:

[phwuil]

Hi @kenneth-lee-ch ,

Is this what you want to do ? (I do not really understand what you mean by"given a string")

Note that when the Potential is on a white background, it means that it is not a CPT anymore :-)

[kenneth-lee-ch]

@phwuil I am just thinking about how to get the conditional probability given any target and conditioning set from a CPT and be able to output a prediction according to the largest probability given a sample. I was thinking to get getPosterior, but it doesn't seem like the same.

[phwuil]

Ok, if I understand correctly this time, you want to use : LazyPropagation.evidenceIImpact :


Note that evidenceImpact above returns P(C|A,F) since E is d-separated from C given A and F => P(C|A,E,F)=P(C|A,F)

[kenneth-lee-ch]

How is this different from getPosterior()?

[phwuil]

It is a getPosterior for all the possible configurations of the variables in the evidence.

getPosterior : P(X |Y=1,Z=0)
evidenceImpact : the whole P(X|Y=y,Z=z) forall y,z

Moreover, it performs this d-separation analysis...

And it takes advantage of aGrUM's ability to make incremental inferences: some computations can be factored. Such that an evidenceImpact if faster than several getPosteriors.

[kenneth-lee-ch]

@phwuil I have a follow-up question. If I use causalImpact to get the post-interventional distribution from the returned argument computation. How can I manipulate the probability table such that the one shown on the screenshot such that the column corresponding to a specific variable I look for e.g. C instead of L? In other words, swapping the column and a specific row according to a specific variable name? I don't know how to do that for any arbitrary table output by causalImpact.

The reason why I want to do that is that I want to get a list of values (0 and 1) according to the max value of a specific variable in the table. For example, in the screenshot, if my specific variable is W, then I would get all 0's as the list of values because 0 always corresponds to the higher probability in this case.

[phwuil]

Hi @kenneth-lee-ch , you have two methods of Potential to do so : p.putFirst(varname) and p.reorganize(list_of_varnames) :

By the way, if I well understood what you would like :

[kenneth-lee-ch]

@phwuil If I have a dataframe with binary variables F, C, L as columns, let's say I want to treat entries of your bestW potential as the predicted value (e.g. when F=0, C=0, L=0, my predicted value is 0; when F=1, C=0, L=0, my predicted value is 1), what's the efficient way to fill in the dataframe with these predicted values accordingly?

[phwuil]

p.topandas()?

[kenneth-lee-ch]

@phwuil No, that's not what I meant. Suppose you have a dataframe with columns E,C,L and they are all binary variables. Now, predict the value of W based on the potential CPT with the entries of E,C,L you have from the dataframe. Let's say the first row of the dataframe tells you L=0, C=0, E= 0, then you will predict W=0 based on the largest probability between W=0 vs W=1

[phwuil]

Given the bestW Potential above, it is pretty straightforward :-)

Just a tip : df.to_dict(orient='records')

... and then : [(d,bestW[d] ) for d in df.to_dict(orient='records')]

[kenneth-lee-ch]

@phwuil I actually had a way before I asked the question. Just wanted to see if there is a more efficient way that can be done through the package functions. Thanks though.