def con_Interval():
X, y, vectorizer = get_X_y()
#BS_PR.sort_ratio(X,y,vectorizer)
n_samples = 1000
bs_indexes = bootstrap_indexes(X,n_samples)
w_lists = np.zeros((n_samples,X.shape[1]))
lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto')
kf = cross_validation.StratifiedKFold(y,n_folds=5,shuffle=True)
parameters = {"C":[100,10,1.0,.1, .01, .001,0.0001]}
clf0 = grid_search.GridSearchCV(lr, parameters,scoring='roc_auc',cv=kf)
clf0.fit(X,y)
best_C = clf0.best_params_['C']
for i in range(n_samples):
train_X = X[bs_indexes[i]]
train_Y = y[bs_indexes[i]]
lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto',C=best_C)
lr.fit(train_X,train_Y)
w = lr.coef_
w_lists[i] = w
print('iteration',i)
mean = np.mean(w_lists,axis=0)
std = np.std(w_lists,axis=0)
p_lower = mean - (1.96)*std
p_upper = mean + (1.96)*std
sort_p_lower = sorted(zip(p_lower.tolist(),vectorizer.get_feature_names(),range(len(mean))),reverse=True)
sort_p_upper = sorted(zip(p_upper.tolist(),vectorizer.get_feature_names(),range(len(mean))))
save_dict = {}
save_dict["w_list"] = w_lists
save_dict["sort_p_lower"] = sort_p_lower
save_dict["sort_p_upper"] = sort_p_upper
save_dict["mean"] = list(mean)
dict_file = open("BS_NConPR/coefficient.pkl","wb")
cPickle.dump(save_dict,dict_file,cPickle.HIGHEST_PROTOCOL)
dict_file.close()
#set break point here
texify_most_informative_features(sort_p_lower,sort_p_upper)
#draw top features for positive instances
for i in range(5):
values = w_lists[:,sort_p_lower[i][2]]
plt.hist(values,bins=20)
plt.title(vectorizer.get_feature_names()[sort_p_lower[i][2]])
plt.xlabel("Value")
plt.ylabel("Frequency")
plt.savefig("BS_NConPR/"+vectorizer.get_feature_names()[sort_p_lower[i][2]]+"_L2.png")
plt.clf()
#draw top features for negative instances
for i in range(5):
plt.clf()
values = w_lists[:,sort_p_upper[i][2]]
plt.hist(values,bins=20)
plt.title(vectorizer.get_feature_names()[sort_p_upper[i][2]])
plt.xlabel("Value")
plt.ylabel("Frequency")
plt.savefig("BS_NConPR/"+vectorizer.get_feature_names()[sort_p_upper[i][2]]+"_L2.png")
plt.clf()
plot_Features(sort_p_lower,sort_p_upper,X,y,vectorizer)
def con_Interval():
X, y, vectorizer = get_X_y()
n_samples = 1000
bs_indexes = bootstrap_indexes(X, n_samples)
w_lists = np.zeros((n_samples, X.shape[1]))
lr = LogisticRegression(penalty="l2",
fit_intercept=True,
class_weight='auto')
kf = cross_validation.StratifiedKFold(y, n_folds=5, shuffle=True)
parameters = {"C": [100, 10, 1.0, .1, .01, .001, 0.0001]}
clf0 = grid_search.GridSearchCV(lr, parameters, scoring='roc_auc', cv=kf)
clf0.fit(X, y)
best_C = clf0.best_params_['C']
print "best AUC score is: " + str(clf0.best_score_)
for i in range(n_samples):
train_X = X[bs_indexes[i]]
train_Y = y[bs_indexes[i]]
#lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto',C=best_C)
#lr.fit(train_X,train_Y)
clf = SGDClassifier(loss="log",
alpha=1.0 / best_C,
n_iter=np.ceil(10**6 / train_X.shape[0]),
class_weight="auto").fit(train_X, train_Y)
w = clf.coef_
w_lists[i] = w
print('iteration', i)
CI_hash_pos = {}
CI_hash_neg = {}
mean = np.mean(w_lists, axis=0)
std = np.std(w_lists, axis=0)
p_lower = mean - (1.96) * std
p_upper = mean + (1.96) * std
sort_p_lower = sorted(zip(p_lower.tolist(), vectorizer.get_feature_names(),
range(len(mean))),
reverse=True)
sort_p_upper = sorted(
zip(p_upper.tolist(), vectorizer.get_feature_names(),
range(len(mean))))
save_dict = {}
save_dict["w_list"] = w_lists
save_dict["sort_p_lower"] = sort_p_lower
save_dict["sort_p_upper"] = sort_p_upper
dict_file = open("reuters/coefficient.pkl", "wb")
cPickle.dump(save_dict, dict_file, cPickle.HIGHEST_PROTOCOL)
dict_file.close()
texify_most_informative_features(sort_p_lower, sort_p_upper)
#draw top features for positive instances
plot_index = [1, 2, 4, 8]
plt.figure(1)
def con_Interval():
X, y, vectorizer = get_X_y()
n_samples = 1000
bs_indexes = bootstrap_indexes(X, n_samples)
w_lists = np.zeros((n_samples, X.shape[1]))
lr = LogisticRegression(penalty="l2", fit_intercept=True, class_weight="auto")
kf = cross_validation.StratifiedKFold(y, n_folds=5, shuffle=True)
parameters = {"C": [100, 10, 1.0, 0.1, 0.01, 0.001, 0.0001]}
clf0 = grid_search.GridSearchCV(lr, parameters, scoring="roc_auc", cv=kf)
clf0.fit(X, y)
best_C = clf0.best_params_["C"]
print "best AUC score is: " + str(clf0.best_score_)
for i in range(n_samples):
train_X = X[bs_indexes[i]]
train_Y = y[bs_indexes[i]]
# lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto',C=best_C)
# lr.fit(train_X,train_Y)
clf = SGDClassifier(
loss="log", alpha=1.0 / best_C, n_iter=np.ceil(10 ** 6 / train_X.shape[0]), class_weight="auto"
).fit(train_X, train_Y)
w = clf.coef_
w_lists[i] = w
print ("iteration", i)
CI_hash_pos = {}
CI_hash_neg = {}
mean = np.mean(w_lists, axis=0)
std = np.std(w_lists, axis=0)
p_lower = mean - (1.96) * std
p_upper = mean + (1.96) * std
sort_p_lower = sorted(zip(p_lower.tolist(), vectorizer.get_feature_names(), range(len(mean))), reverse=True)
sort_p_upper = sorted(zip(p_upper.tolist(), vectorizer.get_feature_names(), range(len(mean))))
save_dict = {}
save_dict["w_list"] = w_lists
save_dict["sort_p_lower"] = sort_p_lower
save_dict["sort_p_upper"] = sort_p_upper
dict_file = open("reuters/coefficient.pkl", "wb")
cPickle.dump(save_dict, dict_file, cPickle.HIGHEST_PROTOCOL)
dict_file.close()
texify_most_informative_features(sort_p_lower, sort_p_upper)
# draw top features for positive instances
plot_index = [1, 2, 4, 8]
plt.figure(1)
def con_Interval():
X, y, vectorizer = get_X_y()
#BS_PR.sort_ratio(X,y,vectorizer)
n_samples = 1000
bs_indexes = bootstrap_indexes(X,n_samples)
w_lists = np.zeros((n_samples,X.shape[1]))
lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto')
kf = cross_validation.StratifiedKFold(y,n_folds=5,shuffle=True)
parameters = {"C":[100,10,1.0,.1, .01, .001,0.0001,0.00001,0.000001,0.0000001]}
clf0 = grid_search.GridSearchCV(lr, parameters,scoring='roc_auc',cv=kf)
clf0.fit(X,y)
best_C = clf0.best_params_['C']
lr1 = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto',C=best_C)
scores = cross_validation.cross_val_score(lr1,X,y,cv=kf,scoring = 'roc_auc')
print "best C is: "+str(best_C)
print scores
for i in range(n_samples):
train_X = X[bs_indexes[i]]
train_Y = y[bs_indexes[i]]
lr = LogisticRegression(penalty="l2", fit_intercept=True,class_weight='auto',C=best_C)
lr.fit(train_X,train_Y)
w = lr.coef_
w_lists[i] = w
print('iteration',i)
mean = np.mean(w_lists,axis=0)
std = np.std(w_lists,axis=0)
p_lower = mean - (1.96)*std
p_upper = mean + (1.96)*std
sort_p_lower = sorted(zip(p_lower.tolist(),vectorizer.get_feature_names(),range(len(mean))),reverse=True)
sort_p_upper = sorted(zip(p_upper.tolist(),vectorizer.get_feature_names(),range(len(mean))))
save_dict = {}
save_dict["w_list"] = w_lists
save_dict["sort_p_lower"] = sort_p_lower
save_dict["sort_p_upper"] = sort_p_upper
save_dict["mean"] = list(mean)
dict_file = open("BS_NConPR/coefficient.pkl","wb")
cPickle.dump(save_dict,dict_file,cPickle.HIGHEST_PROTOCOL)
dict_file.close()
#set break point here
texify_most_informative_features(sort_p_lower,sort_p_upper)
#draw top features for positive instances
for i in range(5):
values = w_lists[:,sort_p_lower[i][2]]
plt.hist(values,bins=20)
plt.title(vectorizer.get_feature_names()[sort_p_lower[i][2]])
plt.xlabel("Value")
plt.ylabel("Frequency")
plt.savefig("BS_NConPR/"+vectorizer.get_feature_names()[sort_p_lower[i][2]]+"_L2.png")
plt.clf()
#draw top features for negative instances
for i in range(5):
plt.clf()
values = w_lists[:,sort_p_upper[i][2]]
plt.hist(values,bins=20)
plt.title(vectorizer.get_feature_names()[sort_p_upper[i][2]])
plt.xlabel("Value")
plt.ylabel("Frequency")
plt.savefig("BS_NConPR/"+vectorizer.get_feature_names()[sort_p_upper[i][2]]+"_L2.png")
plt.clf()
plot_Features(sort_p_lower,sort_p_upper,X,y,vectorizer)
