]
split_types = [["splitting"], ["sampling", 2, 2]]
probs = [.5, .7]
selection_methods = ["DSS", "MSS"]
data = [x, t]
for modeling in modelings:
for split_type, prob, selection_method in zip(split_types, probs,
selection_methods):
DSS_selection = barber_candes_selection.BarberCandesSelection(
data,
modeling,
selection_method,
q=q,
split_type=split_type,
prob=prob).selection()
S_dss = DSS_selection.S
fdr_dss = 100 * utils.FDR(S_dss, true_index)
power_dss = 100 * utils.power(S_dss, true_index)
with open('{}-{}.txt'.format(selection_method, type), 'a') as f:
print('------------{} ({}, {})-------------'.format(
selection_method, modeling["model"], modeling["params"]),
file=f)
print(selection_method + "-" + split_type[0] + "-" + str(prob),
file=f)
print("FDR: " + str(fdr_dss) + "%", file=f)
print("power: " + str(power_dss) + "%", file=f)
VI_stats = ["Diff", "Max"]
optimizations = ["SDP", "samplecorr"]
selection_methods = ["knockoff-MX"]
for selection_method in selection_methods:
for optimization in optimizations:
myknockoff = knockoff_features_construction.Knockoff(x, selection_method, optimization)
knockoff_attrs = myknockoff.knockoff_features()
x, x_tilda= knockoff_attrs.X, knockoff_attrs.X_tilde
data = [x, x_tilda, t]
for modeling in modelings:
for VI_stat in VI_stats:
knockoff_selection = barber_candes_selection.BarberCandesSelection(data, modeling, selection_method,q=q, VI_stat=VI_stat).selection()
S_knock = knockoff_selection.S
fdr_knock = 100*utils.FDR(S_knock, true_index)
power_knock = 100*utils.power(S_knock, true_index)
with open('{}-{}.txt'.format(selection_method, type), 'a') as f:
print('------------Knockoff ({})-------------'.format(modeling["model"]), file=f)
print(selection_method +"-"+ optimization +"-"+ modeling["model"] +"-"+ modeling["params"] +"-"+ VI_stat, file=f)
print("FDR: " +str(fdr_knock) + "%", file=f)
print("power: "+str(power_knock) + "%", file=f)
# print(selection_method +"-"+ optimization +"-"+ modeling["model"] +"-"+ modeling["params"] +"-"+ VI_stat)
f.close()
# adaboost method
selection_method = "DSS"
subsample = np.random.binomial(1, .5, size=n)
train_index = (subsample == 1)
valid_index = (subsample == 0)
w_train = model.fit(x[train_index, :], t[train_index].ravel()).feature_importances_.reshape(-1, 1)
w_valid = model.fit(x[valid_index, :], t[valid_index].ravel()).feature_importances_.reshape(-1, 1)
w_dss = np.vstack((w_train, w_valid))
DSS_selection = barber_candes_selection.BarberCandesSelection(modeling=modeling, selection_method=selection_method, q=q, w=w_dss).selection()
S_dss = DSS_selection.S
fdr_dss = 100*utils.FDR(S_dss, true_index)
power_dss = 100*utils.power(S_dss, true_index)
print('------------{} ({}, {})-------------'.format(selection_method, modeling["model"], modeling["params"]))
print("FDR: " +str(fdr_dss) + "%")
print("power: "+str(power_dss) + "%")
################################################################################################################
############################ knockoff selection
################################################################################################################