def scnn_proportion_experiment(data_fn, name, n_hops, prop_valid, prop_test, transform_fn=util.rw_laplacian, transform_name='rwl'):
print 'Running node experiment (%s)...' % (name,)
A, X, Y = data_fn()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
valid_start = int(n_nodes * (1 - (prop_valid + prop_test)))
test_start = int(n_nodes * (1 - prop_test))
valid_indices = indices[valid_start:test_start]
test_indices = indices[test_start:]
for train_prop in [x / 10.0 for x in range(1, 11)]:
train_end = int(valid_start * train_prop)
train_indices = indices[:train_end]
scnn = SCNN(n_hops=n_hops, transform_fn=transform_fn)
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
probs = scnn.predict_proba(X, test_indices)
print probs
preds = scnn.predict(X, test_indices)
actuals = np.argmax(Y[test_indices,:], axis=1)
f1_micro = f1_score(actuals, preds, average='micro')
f1_macro = f1_score(actuals, preds, average='macro')
accuracy = accuracy_score(actuals, preds)
print 'form: name,n_hops,transform_name,micro_f,macro_f,accuracy'
print '###RESULTS###: %s,%d,%.2f,%s,%.8f,%.8f,%.8f' % (name, n_hops, train_prop, transform_name, f1_micro, f1_macro, accuracy)
def node_experiment(data_fn, name, n_hops, transform_fn=util.rw_laplacian, transform_name='rwl'):
print 'Running node experiment (%s)...' % (name,)
A, X, Y = data_fn()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
np.random.shuffle(indices)
train_indices = indices[:n_nodes // 3]
valid_indices = indices[n_nodes // 3:(2* n_nodes) // 3]
test_indices = indices[(2* n_nodes) // 3:]
scnn = SCNN(n_hops=n_hops, transform_fn=transform_fn)
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
probs = scnn.predict_proba(X, test_indices)
print probs
preds = scnn.predict(X, test_indices)
actuals = np.argmax(Y[test_indices,:], axis=1)
f1_micro = f1_score(actuals, preds, average='micro')
f1_macro = f1_score(actuals, preds, average='macro')
accuracy = accuracy_score(actuals, preds)
print 'form: name,n_hops,transform_name,micro_f,macro_f,accuracy'
print '###RESULTS###: %s,%d,%s,%.8f,%.8f,%.8f' % (name, n_hops, transform_name, f1_micro, f1_macro, accuracy)
def test_scnn():
from sklearn.metrics import f1_score
A, X, Y = data.parse_cora()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
train_indices = indices[:n_nodes // 3]
valid_indices = indices[n_nodes // 3:(2* n_nodes) // 3]
test_indices = indices[(2* n_nodes) // 3:]
scnn = SCNN(n_hops=2)
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
preds = scnn.predict(A, X, test_indices)
actuals = np.argmax(Y[test_indices,:], axis=1)
print 'F score: %.4f' % (f1_score(actuals, preds))
def test_scnn():
from sklearn.metrics import f1_score
A, X, Y = data.parse_cora()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
train_indices = indices[:n_nodes // 3]
valid_indices = indices[n_nodes // 3:(2 * n_nodes) // 3]
test_indices = indices[(2 * n_nodes) // 3:]
scnn = SCNN(n_hops=2)
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
preds = scnn.predict(A, X, test_indices)
actuals = np.argmax(Y[test_indices, :], axis=1)
print 'F score: %.4f' % (f1_score(actuals, preds))
def node_experiment(data_fn,
name,
n_hops,
transform_fn=util.rw_laplacian,
transform_name='rwl'):
print 'Running node experiment (%s)...' % (name, )
A, X, Y = data_fn()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
np.random.shuffle(indices)
train_indices = indices[:n_nodes // 3]
valid_indices = indices[n_nodes // 3:(2 * n_nodes) // 3]
test_indices = indices[(2 * n_nodes) // 3:]
scnn = SCNN(n_hops=n_hops, transform_fn=transform_fn)
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
probs = scnn.predict_proba(X, test_indices)
print probs
preds = scnn.predict(X, test_indices)
actuals = np.argmax(Y[test_indices, :], axis=1)
f1_micro = f1_score(actuals, preds, average='micro')
f1_macro = f1_score(actuals, preds, average='macro')
accuracy = accuracy_score(actuals, preds)
print 'form: name,n_hops,transform_name,micro_f,macro_f,accuracy'
print '###RESULTS###: %s,%d,%s,%.8f,%.8f,%.8f' % (
name, n_hops, transform_name, f1_micro, f1_macro, accuracy)
def scnn_proportion_experiment(data_fn,
name,
n_hops,
prop_valid,
prop_test,
transform_fn=util.rw_laplacian,
transform_name='rwl'):
print 'Running node experiment (%s)...' % (name, )
A, X, Y = data_fn()
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
valid_start = int(n_nodes * (1 - (prop_valid + prop_test)))
test_start = int(n_nodes * (1 - prop_test))
valid_indices = indices[valid_start:test_start]
test_indices = indices[test_start:]
for train_prop in [x / 10.0 for x in range(1, 11)]:
train_end = int(valid_start * train_prop)
train_indices = indices[:train_end]
scnn = SCNN(n_hops=n_hops, transform_fn=transform_fn)
scnn.fit(A,
X,
Y,
train_indices=train_indices,
valid_indices=valid_indices)
probs = scnn.predict_proba(X, test_indices)
print probs
preds = scnn.predict(X, test_indices)
actuals = np.argmax(Y[test_indices, :], axis=1)
f1_micro = f1_score(actuals, preds, average='micro')
f1_macro = f1_score(actuals, preds, average='macro')
accuracy = accuracy_score(actuals, preds)
print 'form: name,n_hops,transform_name,micro_f,macro_f,accuracy'
print '###RESULTS###: %s,scnn%d%s,%.6f,%.8f,%.8f,%.8f' % (
name, n_hops, transform_name, train_prop, f1_micro, f1_macro,
accuracy)
import numpy as np from scnn import SCNN, data from sklearn.metrics import f1_score # Parse the cora dataset and return an adjacency matrix, a design matrix, and a 1-hot label matrix A, X, Y = data.parse_cora() # Construct array indices for the training, validation, and test sets n_nodes = A.shape[0] indices = np.arange(n_nodes) train_indices = indices[:n_nodes // 3] valid_indices = indices[n_nodes // 3:(2* n_nodes) // 3] test_indices = indices[(2* n_nodes) // 3:] # Instantiate an SCNN and fit it to cora scnn = SCNN() scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices) # Predict labels for the test set preds = scnn.predict(A, X, test_indices) actuals = np.argmax(Y[test_indices,:], axis=1) # Display performance print 'F score: %.4f' % (f1_score(actuals, preds))
import numpy as np from scnn import SCNN, data from sklearn.metrics import f1_score # Parse the cora dataset and return an adjacency matrix, a design matrix, and a 1-hot label matrix A, X, Y = data.parse_cora() # Construct array indices for the training, validation, and test sets n_nodes = A.shape[0] indices = np.arange(n_nodes) train_indices = indices[:n_nodes // 3] valid_indices = indices[n_nodes // 3:(2 * n_nodes) // 3] test_indices = indices[(2 * n_nodes) // 3:] # Instantiate an SCNN and fit it to cora scnn = SCNN() scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices) # Predict labels for the test set preds = scnn.predict(A, X, test_indices) actuals = np.argmax(Y[test_indices, :], axis=1) # Display performance print 'F score: %.4f' % (f1_score(actuals, preds))
