def edge_errors(pred, target):
"""
Counts all types of edge errors (false negatives, false positives, reversed edges)
Parameters:
-----------
pred: nx.DiGraph or ndarray
The predicted adjacency matrix
target: nx.DiGraph or ndarray
The true adjacency matrix
Returns:
--------
fn, fp, rev
"""
true_labels = retrieve_adjacency_matrix(target)
predictions = retrieve_adjacency_matrix(
pred,
target.nodes() if isinstance(target, nx.DiGraph) else None)
diff = true_labels - predictions
rev = (((diff + diff.transpose()) == 0) & (diff != 0)).sum() / 2
# Each reversed edge necessarily leads to one fp and one fn so we need to subtract those
fn = (diff == 1).sum() - rev
fp = (diff == -1).sum() - rev
return fn, fp, rev
def edge_accurate(pred, target):
"""
Counts the number of edge in ground truth DAG, true positives and the true
negatives
Parameters:
-----------
pred: nx.DiGraph or ndarray
The predicted adjacency matrix
target: nx.DiGraph or ndarray
The true adjacency matrix
Returns:
--------
total_edges, tp, tn
"""
true_labels = retrieve_adjacency_matrix(target)
predictions = retrieve_adjacency_matrix(pred, target.nodes() if isinstance(target, nx.DiGraph) else None)
total_edges = (true_labels).sum()
tp = ((predictions == 1) & (predictions == true_labels)).sum()
tn = ((predictions == 0) & (predictions == true_labels)).sum()
return total_edges, tp, tn
def generator(mechanism = 'linear'):
number_parents = 20 #int(input('Desired number of parents? '))
number_of_data_points = 5000 #int(input('Desired number of data points? '))
generator = AcyclicGraphGenerator(mechanism, nodes=number_parents, parents_max=3, noise_coeff=.4, npoints=number_of_data_points)
data, graph = generator.generate()
#generator.to_csv('generated_graph')
# Save true matrix
true_matrix = retrieve_adjacency_matrix(graph)
savetxt('true_CM.csv', true_matrix, delimiter=',')
# Normalise the data
n = len(data)
data = np.array(data)
for i in range(len(data.T)):
data[:,i] = (data[:,i]-min(data[:,i]))/(max(data[:,i])-min(data[:,i]))
data = np.concatenate((np.ones(n).reshape(-1,1), data), axis =1)
# Save current version in home folder
savetxt('combined.csv', data, delimiter=',')
run = 0
name = 'ArchivedData/combined_' + str(mechanism) + '_' + str(run) + '.csv'
name_matrix = 'ArchivedData/True_CM_' + str(mechanism) + '_' + str(run) + '.csv'
# Check if the item alreay exists
while os.path.exists(name):
run += 1
name = 'ArchivedData/combined_' + str(mechanism) + '_' + str(run) + '.csv'
# Save
savetxt(name_matrix, true_matrix, delimiter=',')
savetxt(name, data, delimiter=',')
savetxt(name, data, delimiter=',')
# Generate Graph
mechanism = str(input('Which mechanism do you select to generator the data? '))
number_parents = 20 #int(input('Desired number of parents? '))
number_of_data_points = int(input('Desired number of data points? '))
generator = AcyclicGraphGenerator(mechanism, nodes=number_parents, parents_max=3, noise_coeff=.4, npoints=number_of_data_points)
data, graph = generator.generate()
#generator.to_csv('generated_graph')
# Save true matrix
true_matrix = retrieve_adjacency_matrix(graph)
savetxt('true_CM.csv', true_matrix, delimiter=',')
plt.figure(figsize=(40,20))
nx.draw_networkx(graph, font_size=10) # The plot function allows for quick visualization of the graph.
plt.show()
# Normalise the data
n = len(data)
data = np.array(data)
for i in range(len(data.T)):
data[:,i] = (data[:,i]-min(data[:,i]))/(max(data[:,i])-min(data[:,i]))
def baselines(data):
# Tests
start_time = time.time()
obj = PC()
output = obj.predict(data)
adj_mat = nx.adjacency_matrix(output).todense()
output = clr(adj_mat)
output[np.isnan(output)] = 0
output[np.isposinf(output)] = 1
predicted = retrieve_adjacency_matrix(output)
true_matrix = pd.read_csv('true_CM.csv', header=None)
true_matrix = np.array(true_matrix)
shd = SHD(np.array(true_matrix), predicted, double_for_anticausal=False)
aupr, curve = precision_recall(np.array(true_matrix), output)
end_time = (time.time() - start_time)
print("--- Execution time : %4.4s seconds ---" % end_time)
results_pc = ['PC', aupr, shd, end_time]
print(results_pc)
# Tests
start_time = time.time()
obj = GES()
output = obj.predict(data)
adj_mat = nx.adjacency_matrix(output).todense()
output = clr(adj_mat)
output[np.isnan(output)] = 0
output[np.isposinf(output)] = 1
predicted = retrieve_adjacency_matrix(output)
true_matrix = pd.read_csv('true_CM.csv', header=None)
true_matrix = np.array(true_matrix)
shd = SHD(np.array(true_matrix), predicted, double_for_anticausal=False)
aupr, curve = precision_recall(np.array(true_matrix), output)
end_time = (time.time() - start_time)
print("--- Execution time : %4.4s seconds ---" % end_time)
results_ges = ['GES', aupr, shd, end_time]
print(results_ges)
# Tests
start_time = time.time()
obj = LiNGAM()
output = obj.predict(data)
adj_mat = nx.adjacency_matrix(output).todense()
output = clr(adj_mat)
output[np.isnan(output)] = 0
output[np.isposinf(output)] = 1
predicted = retrieve_adjacency_matrix(output)
true_matrix = pd.read_csv('true_CM.csv', header=None)
true_matrix = np.array(true_matrix)
shd = SHD(np.array(true_matrix), predicted, double_for_anticausal=False)
aupr, curve = precision_recall(np.array(true_matrix), output)
end_time = (time.time() - start_time)
print("--- Execution time : %4.4s seconds ---" % end_time)
results_lingam = ['LiNGAM', aupr, shd, end_time]
print(results_lingam)
# Tests
start_time = time.time()
obj = CCDr()
output = obj.predict(data)
adj_mat = nx.adjacency_matrix(output).todense()
output = clr(adj_mat)
output[np.isnan(output)] = 0
output[np.isposinf(output)] = 1
predicted = retrieve_adjacency_matrix(output)
true_matrix = pd.read_csv('true_CM.csv', header=None)
true_matrix = np.array(true_matrix)
shd = SHD(np.array(true_matrix), predicted, double_for_anticausal=False)
aupr, curve = precision_recall(np.array(true_matrix), output)
end_time = (time.time() - start_time)
print("--- Execution time : %4.4s seconds ---" % end_time)
results_ccdr = ['CCDR', aupr, shd, end_time]
print(results_ccdr)
return results_pc, results_ges, results_lingam, results_ccdr
results_ccdr = ['CCDR', aupr, shd, end_time]
print(results_ccdr)
return results_pc, results_ges, results_lingam, results_ccdr
# Tests
start_time = time.time()
obj = CGNN(nruns=1, train_epochs=500, test_epochs=500)
output = obj.predict(data)
adj_mat = nx.adjacency_matrix(output).todense()
output = clr(adj_mat)
output[np.isnan(output)] = 0
output[np.isposinf(output)] = 1
predicted = retrieve_adjacency_matrix(output)
true_matrix = pd.read_csv('true_CM.csv', header=None)
true_matrix = np.array(true_matrix)
shd = SHD(np.array(true_matrix), predicted, double_for_anticausal=False)
aupr, curve = precision_recall(np.array(true_matrix), output)
end_time = (time.time() - start_time)
print("--- Execution time : %4.4s seconds ---" % end_time)
results_cgnn = ['CGNN', aupr, shd, end_time]
print(results_cgnn)