def i_EM(P,
U,
max_imbalance=10.0,
max_pos_ratio=1.0,
tolerance=0.1,
verbose=False):
"""all-in-one PU method: I-EM algorithm for positive set P and unlabelled set U
iterate NB classifier with updated labels for unlabelled set (initially negative) until convergence
if U is much larger than P, randomly samples max_imbalance*|P| docs from U"""
print("Running I-EM")
# if num_rows(U) > max_imbalance * num_rows(P):
# U = np.array(random.sample(list(U), int(max_imbalance * num_rows(P))))
model = iterate_EM(P,
U,
tolerance=tolerance,
max_pos_ratio=max_pos_ratio,
clf_selection=False,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def biased_SVM_grid_search(P,
U,
Cs=None,
kernel='linear',
n_estimators=9,
verbose=False):
if Cs is None:
Cs = [10**x for x in range(-12, 12, 2)]
if verbose:
print(
"Running Biased-SVM with balanced class weights and grid search over",
len(Cs), "C values")
model = BaggingClassifier(LinearSVC())
grid_search = GridSearchCV(
model,
param_grid={
'base_estimator__C': Cs,
'base_estimator__class_weight': ['balanced'],
### not applicable for LinearSVC
# 'base_estimator__kernel' : [kernel],
# 'base_estimator__cache_size' : [8000],
# 'base_estimator__probability' : [True],
### fit parameters for Bagging wrapper
'bootstrap': [True],
'n_estimators': [n_estimators],
### parallelization incompatible with multiprocessing
# 'n_jobs' : [n_estimators]
},
scoring=pu_scorer,
verbose=0)
if verbose:
print("Grid searching parameters for biased-SVM")
X = concatenate((P, U))
y = concatenate((ones(num_rows(P)), zeros(num_rows(U))))
grid_search.fit(X, y)
if verbose:
train_report(grid_search.best_estimator_, P, U)
print("Biased-SVM parameters:", grid_search.best_params_, "\tPU score:",
grid_search.best_score_)
return grid_search.best_estimator_
def standalone_rocchio(P, U, alpha=16, beta=4, verbose=False):
"""1-step Rocchio method"""
print("Running Rocchio")
if verbose:
print("Building Rocchio model to determine Reliable Negative examples")
model = rocchio(P, U, alpha=alpha, beta=beta)
y_U = model.predict(U)
U_minus_RN, RN = partition_pos_neg(U, y_U)
if verbose:
print("Reliable Negative examples in U:", num_rows(RN), "(",
100 * num_rows(RN) / num_rows(U), "%)")
train_report(model, P, U)
return model
def cr_SVM(P,
U,
max_neg_ratio=0.1,
noise_lvl=0.2,
alpha=16,
beta=4,
kernel=None,
C=0.1,
verbose=False):
"""Two-Step technique based on Cosine Similarity, Rocchio and SVM
Step 1.1: Find Potentially Negative docs (less similar to mean(P) than noise_lvl of docs in P)
Step 1.2: Find Reliable Negative docs using Rocchio (similarity to mean positive/PN vector)
Step 2: Iterate SVM starting from P and RN sets until classification of U converges
noise level is quite crucial, should be >=20% to give reasonable results"""
print("Running CR-SVM")
# step 1
if verbose:
print("Determining RN using Cosine Similarity threshold and Rocchio\n")
U_minus_RN, RN = get_RN_cosine_rocchio(P,
U,
noise_lvl=noise_lvl,
alpha=alpha,
beta=beta,
verbose=verbose)
# step2
if verbose:
print("\nIterating SVM with P, U-RN, and RN")
model = iterate_SVM(P,
U_minus_RN,
RN,
kernel=kernel,
C=C,
max_neg_ratio=max_neg_ratio,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def s_EM(P,
U,
spy_ratio=0.1,
max_pos_ratio=1.0,
tolerance=0.1,
noise_lvl=0.1,
clf_selection=True,
verbose=False):
"""S-EM two-step PU learning as described in \"Partially Supervised Classification...\".
1st step: get Reliable Negative documents using Spy Documents
2nd step: iterate EM with P, U-RN, and RN
"""
print("Running S-EM")
# P, U = arrays([P, U]
# step 1
if verbose:
print(
"Determining confidence threshold using Spy Documents and I-EM\n")
U_minus_RN, RN = get_RN_Spy_Docs(P,
U,
spy_ratio=spy_ratio,
tolerance=tolerance,
noise_lvl=noise_lvl,
verbose=verbose)
# step2
if verbose:
print("\nIterating I-EM with P, U-RN, and RN")
model = run_EM_with_RN(P,
U_minus_RN,
RN,
tolerance=tolerance,
max_pos_ratio=max_pos_ratio,
clf_selection=clf_selection,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def roc_EM(P,
U,
max_pos_ratio=0.5,
tolerance=0.1,
clf_selection=True,
alpha=16,
beta=4,
verbose=False):
"""S-EM two-step PU learning as described in \"Partially Supervised Classification...\".
1st step: get Reliable Negative documents using Spy Documents
2nd step: iterate EM with P, U-RN, and RN
"""
print("Running Roc-EM")
# step 1
if verbose:
print("Determining RN using Rocchio method\n")
U_minus_RN, RN = get_RN_rocchio(P,
U,
alpha=alpha,
beta=beta,
verbose=verbose)
# step2
if verbose:
print("\nIterating I-EM with P, U-RN, and RN")
model = run_EM_with_RN(P,
U_minus_RN,
RN,
tolerance=tolerance,
max_pos_ratio=max_pos_ratio,
clf_selection=clf_selection,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def roc_SVM(P,
U,
max_neg_ratio=0.1,
alpha=16,
beta=4,
kernel=None,
C=0.1,
verbose=False):
"""Two-Step technique based on Rocchio and SVM
Step 1: Find Reliable Negative docs using Rocchio (similarity to mean positive/unlabelled vector)
Step 2: Iterate SVM starting from P and RN sets until classification of U converges"""
print("Running Roc-SVM")
# step 1
if verbose:
print("Determining RN using Rocchio method\n")
U_minus_RN, RN = get_RN_rocchio(P,
U,
alpha=alpha,
beta=beta,
verbose=verbose)
# step2
if verbose:
print("\nIterating SVM with P, U-RN, and RN")
model = iterate_SVM(P,
U_minus_RN,
RN,
kernel=kernel,
C=C,
max_neg_ratio=max_neg_ratio,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def spy_SVM(P,
U,
spy_ratio=0.1,
max_neg_ratio=0.1,
tolerance=0.1,
noise_lvl=0.1,
verbose=False):
"""S-EM two-step PU learning as described in \"Partially Supervised Classification...\".
1st step: get Reliable Negative documents using Spy Documents
2nd step: iterate EM with P, U-RN, and RN
"""
print("Running Spy-SVM")
# step 1
if verbose:
print(
"Determining confidence threshold using Spy Documents and I-EM\n")
U_minus_RN, RN = get_RN_Spy_Docs(P,
U,
spy_ratio=spy_ratio,
tolerance=tolerance,
noise_lvl=noise_lvl,
verbose=verbose)
# step2
if verbose:
print("\nIterating SVM with P, U-RN, and RN")
model = iterate_SVM(P,
U_minus_RN,
RN,
max_neg_ratio=max_neg_ratio,
verbose=verbose)
if verbose:
train_report(model, P, U)
return model
def biased_SVM_weight_selection(P,
U,
Cs_neg=None,
Cs_pos_factors=None,
Cs=None,
kernel='linear',
test_size=0.2,
verbose=False):
"""run biased SVMs with combinations of class weight values, choose the one with the best pu_measure"""
# default values
if Cs is None:
Cs = [10**x for x in range(-12, 12, 2)]
if Cs_neg is None:
Cs_neg = [1] # arange(0.01, 0.63, 0.04)
if Cs_pos_factors is None:
Cs_pos_factors = range(1, 1100, 200)
Cs = [(C, C_neg * j, C_neg) for C in Cs for C_neg in Cs_neg
for j in Cs_pos_factors]
if verbose:
print(
"Running Biased-SVM with range of C and positive class weight factors.",
num_rows(Cs), "parameter combinations.")
P_train, P_test = train_test_split(P, test_size=test_size)
U_train, U_test = train_test_split(U, test_size=test_size)
X = concatenate((P_train, U_train))
y = concatenate((ones(num_rows(P_train)), zeros(num_rows(U_train))))
# with Pool(processes=min(cpu_count() - 1, num_rows(Cs))) as p:
score_weights = map(
partial(eval_params,
X_train=X,
y_train=y,
P_test=P_test,
U_test=U_test,
kernel=kernel), Cs)
best_score_params = max(score_weights, key=lambda tup: tup[0])
[print(s) for s in score_weights]
if verbose:
print("\nBest model has parameters", best_score_params[1],
"and PU-score", best_score_params[0])
print("Building final classifier")
model = build_biased_SVM(concatenate((P, U)),
concatenate(
(ones(num_rows(P)), zeros(num_rows(U)))),
C_pos=best_score_params[1]['C_pos'],
C_neg=best_score_params[1]['C_neg'],
C=best_score_params[1]['C'],
probability=True,
kernel=kernel)
if verbose:
train_report(model, P, U)
print("Returning Biased-SVM with parameters", best_score_params[1],
"and PU-score", best_score_params[0])
return model