random BN specifications

[kenneth-lee-ch]

Hi, is there a way to randomly generate BN with the ability to make half of the nodes binary and half of the nodes have 20 states for example? or picking uniformly at random to assign different numbers of states on top of randomBN ?

[phwuil]

Hi @kenneth-lee-ch ,

both gum.BNGenerator and the shortcuts gum.randomBN can assign different numbers of states. The domain_size you give is just the max.

and

[kenneth-lee-ch]

@phwuil Thank you so much! I love the function name. You are the best.

[kenneth-lee-ch]

@phwuil I tried to generate DAGs that have 80% of the variables as binary variables and 20% of them to have 80 states, I set the number of variables to be 15 with 70 edges. I often ran into the following memory errors, do you know why? It was working when I :
'''
return _pyAgrum.BayesNet_endTopologyTransformation(self)
Exception: std::bad_alloc
'''

Here is the code I used:

def randomBNwithSpecificStates(nodes:int,arcs:int, states: List[int], p:float, graphidx):
    table = {}
    gum.initRandom(graphidx)
    g=gum.BNGenerator()
    tmp=g.generate(nodes,arcs,2)
    bn=gum.BayesNet()
    # Nodes
    v=list(tmp.names())
    random.shuffle(v)
    # h=len(v)//2
    for name in v:
        s = np.random.choice(a=np.array(states), size=1, p=p)
        state_num = s[0]
        bn.add(name, int(state_num))
        id = bn.ids(set([name]))
        table[id[0]] = state_num
    
    # arcs
    bn.beginTopologyTransformation()
    for a,b in tmp.arcs():
        bn.addArc(tmp.variable(a).name(),tmp.variable(b).name())
    bn.endTopologyTransformation()
    bn.generateCPTs()
    # output_dict = {value: key for key, value in table.items()}
    print(table)
    return bn, table

n = 30
num_edges = 70
for i in range(100):
       randomBNwithSpecificStates(n, num_edges, states =[2, 80], p=[0.2, 0.8], graphidx=i)

[phwuil]

The moment where you endTopologyTransformation is the moment where all the CPTs are created. It may happen that when you created your graph, unfortunately, a 80-states variable is the children of many 80-state other variables. This leads to a massive amount of memory needed for this CPT.

• 3 80-states parents for an 80-state variable means 8*80^4 = 312 Mo for this variable only.
• 4 80-states parents for an 80-state variable is more than 20Go for this variable only.

There is nothing to do about it : randomBNwithSpecificStates may lead to BN bigger than your memory. You have to change this algorithm... by restricting the number of parents (especially) with 80-states, for instance.

[kenneth-lee-ch]

@phwuil do you know how to restrict the number of parents in this case?

[phwuil]

yes : you have to modify tmp before creating bn in your code : fon instance, each time a node has too many parents, randomly remove a parent.