Mutual Information in Pgmpy

bn_mi.py

# @param infer: a pgmpy Inference object over a BayesianNetwork (e.g. VariableElimination)
# @param node: a string identifier for one of the nodes in the BN

def mutual_info_with_percents(infer, node):
    nodes = list(infer.model.nodes())

    mis = []
    mi_percentages = []
    for other_node in nodes:
        mis.append(mi(infer, node, other_node))
        mi_percentages.append(mi_percentchange(infer, node, other_node))

    return pd.DataFrame({"MutualInfo":mis, "MIPercentChange":mi_percentages}, index=nodes)
    

def mi(infer, node1, node2):
    if node1 == node2:
        return np.nan
    proby1 = infer.query([node1])
    proby2 = infer.query([node2])    
    proby = infer.query([node1, node2], joint=True)
    probs = np.outer(proby1.values, proby2.values)

    return (np.sum(proby.values*np.log(proby.values) - proby.values*np.log(probs)))


def mi_percentchange(infer, node1, node2):
    if node1 == node2:
        return np.nan
    
    proby1 = infer.query([node1])
    proby2 = infer.query([node2])    
    proby = infer.query([node1, node2], joint=True)

    probs = np.outer(proby1.values, proby2.values)

    
    cond_entropy = np.sum(proby.values*np.log(probs))
    entropy = np.sum(proby.values*np.log(proby.values))
    return((1 - cond_entropy/entropy) * -100)