Efficient retrieval of a particular value in CPT given the values

[kenneth-lee-ch]

If I have a sample of values that I want to look up from the bn joint CPT table, is there a way to retrieve it efficiently without storing the entire joint table in pyAgrum?

[kenneth-lee-ch]

I am currently doing this, but it's memory-intensive and it's very slow...

for v in list(nx_dag.nodes):
        v_name = str(v)
        Pa_v = set(nx_dag.predecessors(v))
        Pa_v_names = nodeset_to_nameset(Pa_v)
        P_v_given_Pa_v = bn.evidenceImpact(v_name, list(Pa_v_names))
        int_joint_prob = int_joint_prob * P_v_given_Pa_v
p = int_joint_prob[sample_x.iloc[0].to_dict()]

[phwuil]

Hi again :-),

I have a snippet to propose using bn.jointProbability.

but I have to add some comments :

• it could be more direct (remove the need of a completeInstantiation). We will add this. Thanks :-)

• Note the 2 last dictionaries : For all variable, if you give an int as a value, it is the position of the value in the variable labels (the idLabel). If you want to give the label, even if it is an int (such as 5), you have to give it as a string. Hence, the 2 last dictionaries address the same value in the joint probability.

[kenneth-lee-ch]

@phwuil Thank you so much! Sorry to ask another question...if I have a new sample that I want to use it to update this cpt, what should I do?

[kenneth-lee-ch]

Actually, I wanted to see how to do the following task using pyagrum: if I want to construct the CPT from some datasets and then supposed there is another observation, I want to update my cpt with that new sample, then use the CPT to estimate a probability for that single observation. Currently, I use learner = gum.BNLearner(df_mixed_for_iteration, bn) every time for a newly constructed dataframe by concatenating the dataframe with the new sample, but I wonder if there is an more efficient way to do that.