def test_spline_regression():
mdl = gam.GAM('normal', name='test_additive_regression')
mdl.add_feature(name='hft', type='spline', rel_dof=9.)
X, y = generate_spline_data(1000)
mdl.fit(X, y, verbose=False, plot_convergence=True)
mdl.plot('hft', true_fn=lambda x: np.sin(12. * (x + 0.2)) / (x + 0.2))
def test_linear_regression():
mdl = gam.GAM('normal', name='test_linear_regression')
mdl.add_feature(name='purchases', type='linear', transform=np.log1p)
mdl.add_feature(name='gender', type='categorical')
mdl.add_feature(name='country', type='categorical')
X, y = generate_data(1000)
mdl.fit(X, y, verbose=False, plot_convergence=True)
mdl.summary()
Xtest, ytest = generate_data(100)
yhat = mdl.predict(Xtest)
err = ytest - yhat
print 'MSE:', err.dot(err) / len(err) # MSE
def test_additive_regression():
mdl = gam.GAM('normal', name='test_additive_regression')
mdl.add_feature(name='hft', type='spline', rel_dof=9.)
mdl.add_feature(name='purchases', type='linear', transform=np.log1p)
mdl.add_feature(name='gender', type='categorical')
mdl.add_feature(name='country', type='categorical')
X, y = generate_data(1000, include_hft=True)
mdl.fit(X, y, verbose=False, plot_convergence=True)
mdl.plot('hft', true_fn=gmu_hft)
mdl.summary()
Xtest, ytest = generate_data(100, include_hft=True)
yhat = mdl.predict(Xtest)
err = ytest - yhat
print 'MSE:', err.dot(err) / len(err) # MSE
def test_logistic_regression():
mdl = gam.GAM('binomial', name='test_logistic_regression')
mdl.add_feature(name='purchases', type='linear', transform=np.log1p)
mdl.add_feature(name='gender', type='categorical')
mdl.add_feature(name='country', type='categorical')
X, y = generate_data(1000, link=_logit_link, family=_binomial_family)
mdl.fit(X, y, verbose=False, plot_convergence=True)
mdl.summary()
# Get the "true" probabilities, ytest
Xtest, mu_test = generate_data(100,
link=_logit_link,
family=_binomial_family,
return_mean=True)
mu_hat = mdl.predict(Xtest)
err = mu_test - mu_hat
print 'MSE:', err.dot(err) / len(err) # MSE
def test_cross_validation():
mdl = gam.GAM('normal', name='test_additive_regression')
mdl.add_feature(name='hft', type='spline', rel_dof=9.)
num_training_examples = 1000
X, y = generate_spline_data(num_training_examples)
# Use K-fold cross validation to estimate the optimal smoothing parameter
K = 5
ii = np.random.permutation(num_training_examples)
num_smooths = 20
dev = np.zeros((num_smooths, ))
smoothing = np.linspace(0.5, 5.0, num_smooths)
for j in range(num_smooths):
for i in range(K):
ia = int(i * float(num_training_examples) / K)
ib = int((i + 1) * float(num_training_examples) / K) - 1
traini = np.append(ii[0:ia], ii[ib:num_training_examples])
testi = ii[ia:ib]
Xtraini = X.iloc[traini, :]
ytraini = y[traini]
Xtesti = X.iloc[testi, :]
ytesti = y[testi]
mdl.fit(Xtraini, ytraini, smoothing=smoothing[j])
dev[j] += mdl.deviance(Xtesti, ytesti) / np.size(ytesti)
dev[j] /= K
# Refit model using entire training set and best smoothing parameter
best_smoothing = np.argmin(dev)
mdl.fit(X, y, smoothing=smoothing[best_smoothing])
mdl.plot('hft', true_fn=gmu_hft)
mdl.summary()
Xtest, ytest = generate_spline_data(100)
yhat = mdl.predict(Xtest)
err = ytest - yhat
print 'MSE:', err.dot(err) / len(err)