def knowsim_experiment(scope,
scope_name,
type_list,
count,
newLabels,
tau=1,
kNeighbors=10,
label_num=5):
split_path = 'data/local/split/' + scope_name + '/'
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
repeats = 50
tf_param = {'word': True, 'entity': False, 'we_weight': 0.1}
X_word, newIds, entityIds = GraphGenerator.getTFVectorX(hin, tf_param)
n = X_word.shape[0]
knowsim = sparse.lil_matrix((n, n))
for t in type_list:
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
X_typed, newIds, entityIds = GraphGenerator.getTFVectorX(
hin, tf_param, t)
# make similarity graph
cosX = cosine_similarity(X_typed)
graph = sparse.lil_matrix((n, n))
for i in range(n):
for j in np.argpartition(cosX[i], -kNeighbors)[-kNeighbors:]:
if j == i:
continue
graph[i, j] = cosX[i, j] # np.exp(- (1 - cosX[i, j]) / 0.03) #
graph[j, i] = cosX[i, j] # np.exp(- (1 - cosX[i, j]) / 0.03) #
# calculate laplacian scores
row_sum = graph.sum(axis=1)
laplacian_score = generate_laplacian_score(row_sum, X_word, kNeighbors)
# add meta-path-based similarity to the knowsim
knowsim = knowsim + np.exp(-tau * laplacian_score) * graph
knowsim = knowsim.tocsr()
print 'running lp'
lp_param = {'alpha': 0.98, 'normalization_factor': 5}
ssl = SSLClassifier(knowsim,
newLabels,
scope,
lp_param,
repeatTimes=50,
trainNumbers=label_num,
classCount=count)
ssl.repeatedFixedExperimentwithNewIds(pathPrefix=split_path + 'lb' +
str(label_num).zfill(3) + '_',
newIds=newIds)
return ssl.get_mean()
def run_semihin():
# 20NG
for i in range(2):
scope_name = ng20_scope_names[i]
scope = ng20_scopes[i]
count = ng20_counts[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
newLabels = GraphGenerator.getNewLabels(hin)
tf_param = {'word': True, 'entity': False, 'we_weight': 0.112}
X, newIds, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
print scope_name + ' semihin'
result[i, 7] = semihin_experiment(scope, scope_name, count, X, newIds)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.112}
X, newIds, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
print scope_name + ' semihin+entity'
result[i, 8] = semihin_experiment(scope, scope_name, count, X, newIds)
# GCAT
for i in range(2):
scope_name = gcat_scope_names[i]
scope = gcat_scopes[i]
count = gcat_counts[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
newLabels = GraphGenerator.getNewLabels(hin)
tf_param = {'word': True, 'entity': False, 'we_weight': 0.112}
X, newIds, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
print scope_name + ' semihin'
result[i + 2, 6] = semihin_experiment(scope, scope_name, count, X,
newIds)
with open('data/local/laplacian/' + scope_name + '.x') as f:
X = pk.load(f)
print scope_name + ' semihin+entity'
result[i + 2, 7] = semihin_experiment(scope, scope_name, count, X,
newIds)
def run_svm():
# 20NG
for i in range(2):
scope_name = ng20_scope_names[i]
scope = ng20_scopes[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
print scope_name + ' svm'
tf_param = {'word': True, 'entity': False, 'we_weight': 0.1}
X, doc_new_ids, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
y = GraphGenerator.gety(hin)
result[i, 2] = svm_experiment(scope_name, X, y)
print scope_name + ' svm+entity'
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
X, doc_new_ids, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
y = GraphGenerator.gety(hin)
result[i, 3] = svm_experiment(scope_name, X, y)
# GCAT
for i in range(2):
scope_name = gcat_scope_names[i]
scope = gcat_scopes[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
print scope_name + ' svm'
tf_param = {'word': True, 'entity': False, 'we_weight': 0.1}
X, doc_new_ids, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
y = GraphGenerator.gety(hin)
result[i + 2, 2] = svm_experiment(scope_name, X, y)
print scope_name + ' svm+entity'
with open('data/local/laplacian/' + scope_name + '.x') as f:
X = pk.load(f)
y = GraphGenerator.gety(hin)
result[i + 2, 3] = svm_experiment(scope_name, X, y)
def run_generate_laplacian_score():
print 'generate laplacian score for feature reweighting'
# 20NG
for i in range(2):
scope_name = ng20_scope_names[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
X_word, newIds, entityIds = GraphGenerator.getTFVectorX(hin, tf_param)
for t in NG20TypeList:
tf_param = {'word': True, 'entity': True, 'we_weight': 0.112}
X_typed, newIds, entityIds = GraphGenerator.getTFVectorX(
hin, tf_param, t)
laplacian_score = generate_laplacian_score_vector(
X_typed, X_word, 100)
if not os.path.exists('data/local/laplacian/' + scope_name):
os.makedirs('data/local/laplacian/' + scope_name)
with open(
'data/local/laplacian/' + scope_name + '/' + str(t) +
'_scores', 'w') as f:
pk.dump(laplacian_score, f)
# GCAT
for i in range(2):
scope_name = gcat_scope_names[i]
type_list = GCATTypeList[i]
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
X_word, newIds, entityIds = GraphGenerator.getTFVectorX(hin, tf_param)
for t in type_list:
tf_param = {'word': True, 'entity': True, 'we_weight': 0.112}
X_typed, newIds, entityIds = GraphGenerator.getTFVectorX(
hin, tf_param, t)
laplacian_score = generate_laplacian_score_vector(
X_typed, X_word, 100)
if not os.path.exists('data/local/laplacian/' + scope_name):
os.makedirs('data/local/laplacian/' + scope_name)
with open(
'data/local/laplacian/' + scope_name + '/' + str(t) +
'_scores', 'w') as f:
pk.dump(laplacian_score, f)
def generate_meta_graph(scope, scope_name, type_list, count):
split_path = 'data/local/split/' + scope_name + '/'
pred_path = 'data/local/metagraph/' + scope_name + '/'
if not os.path.exists('data/local/metagraph/' + scope_name + '/'):
os.makedirs('data/local/metagraph/' + scope_name + '/')
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
for t in type_list:
#print t
X, newIds, entitynewIds = GraphGenerator.getTFVectorX(hin, tf_param, t)
n = X.shape[0]
e = X.shape[1]
with open('data/local/laplacian/' + scope_name + '/' + str(t) +
'_scores') as f:
laplacian_score = pk.load(f)
laplacian_score = 20 * np.exp(-laplacian_score * 0.01)
D = sparse.diags(laplacian_score)
X = X * D
X = X.toarray()
graph = np.zeros((n + e, n + e))
graph[0:n, n:n + e] = X
graph[n:n + e, 0:n] = X.transpose()
graph = sparse.csc_matrix(graph)
newLabel = GraphGenerator.getNewLabels(hin)
lp_param = {'alpha': 0.98, 'normalization_factor': 5}
# 3-class classification
lp_candidate = [5]
for lp in lp_candidate:
ssl = SSLClassifier(graph,
newLabel,
scope,
lp_param,
repeatTimes=50,
trainNumbers=lp,
classCount=count)
if not os.path.exists(pred_path + str(t) + '/'):
os.makedirs(pred_path + str(t) + '/')
ssl.repeatedFixedExperimentwithNewIds(
pathPrefix=split_path + 'lb' + str(lp).zfill(3) + '_',
newIds=newIds,
saveProb=True,
savePathPrefix=pred_path + str(t) + '/' + 'lb' +
str(lp).zfill(3))
def ensemble_cotrain_experiment(scope,
scope_name,
type_list,
threshold,
weight,
count,
label_num=5):
pred_path = 'data/local/cotrain/' + scope_name + '/'
split_path = 'data/local/split/' + scope_name + '/'
if not os.path.exists(pred_path):
os.makedirs(pred_path)
with open('data/local/' + scope_name + '.dmp') as f:
hin = pk.load(f)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.1}
c = len(scope)
lb_cand = [label_num]
repeats = 50
# rounds for alternating optimization
rounds = 2
best_res = 0
X_s = {}
tf_param = {'word': True, 'entity': False, 'we_weight': 0.112}
X_word, newIds, entity_new_ids = GraphGenerator.getTFVectorX(
hin, param=tf_param, entity_types=None)
for t in type_list:
if not os.path.exists(pred_path + str(t) + '/'):
os.makedirs(pred_path + str(t) + '/')
with open('data/local/laplacian/' + scope_name + '/' + str(t) +
'_scores') as f:
laplacian_score = pk.load(f)
tf_param = {'word': True, 'entity': True, 'we_weight': 0.112}
X_typed, newIds, entityIds = GraphGenerator.getTFVectorX(
hin, tf_param, t)
laplacian_score = 20 * np.exp(-laplacian_score * 0.01)
# laplacian_score = laplacian_score / np.sum(laplacian_score) * laplacian_score.shape[0]
D = sparse.diags(laplacian_score)
X_typed = X_typed * D
X_s[str(t)] = X_typed
for rd in range(rounds):
round_best_res = 0
round_best_t = ''
# step 1:
# generate output of each meta-path
for t in type_list:
X = X_s[str(t)].toarray()
n = X.shape[0]
e = X.shape[1]
graph = np.zeros((n + e, n + e))
graph[0:n, n:n + e] = X
graph[n:n + e, 0:n] = X.transpose()
graph = sparse.csc_matrix(graph)
newLabel = GraphGenerator.getNewLabels(hin)
lp_param = {
'alpha': 0.98,
'normalization_factor': 5,
'method': 'variant'
}
lb = label_num
ssl = SSLClassifier(graph,
newLabel,
scope,
lp_param,
repeatTimes=repeats,
trainNumbers=lb,
classCount=count)
if rd == 0:
ssl.repeatedFixedExperimentwithNewIds(
pathPrefix=split_path + 'lb' + str(lb).zfill(3) + '_',
newIds=newIds,
saveProb=True,
savePathPrefix=pred_path + str(t) + '/lb' +
str(lb).zfill(3))
else:
inputPredPath = 'data/local/cotrain/' + scope_name + '/lb' + str(
lb).zfill(3) + '_pred_rd_' + str(rd - 1).zfill(3)
ssl.repeatedFixedExpeimentwithInput(
pathPrefix=split_path + 'lb' + str(lb).zfill(3) + '_',
newIds=newIds,
saveProb=True,
savePathPrefix=pred_path + 'lb' + str(lb).zfill(3) + '_' +
str(t),
inputPredPath=inputPredPath)
res = ssl.get_mean()
if res > best_res:
best_res = res
best_t = t
if res > round_best_res:
round_best_res = res
round_best_t = t
print 'Round %d\t%.4f\t%s' % (rd, round_best_res, str(round_best_t))
# step 2:
# propagate pseudo-label for other path
for lb in lb_cand:
results = []
for r in range(repeats):
with open('data/local/split/' + scope_name + '/lb' +
str(lb).zfill(3) + '_' + str(r).zfill(3) +
'_train') as f:
trainLabel = pk.load(f)
with open('data/local/split/' + scope_name + '/lb' +
str(lb).zfill(3) + '_' + str(r).zfill(3) +
'_test') as f:
testLabel = pk.load(f)
numTrain = len(trainLabel)
numTest = len(testLabel)
n = numTrain + numTest
# write output probability
outPred = np.zeros((n, c))
for t in type_list:
typePred = np.zeros((n, c))
with open(pred_path + str(t) + '/lb' + str(lb).zfill(3) +
'_' + str(r).zfill(3) + '_train') as f:
trainPred = pk.load(f)
for i, k in enumerate(trainLabel.keys()):
typePred[k, :] = trainPred[i, :]
with open(pred_path + str(t) + '/lb' + str(lb).zfill(3) +
'_' + str(r).zfill(3) + '_test') as f:
testPred = pk.load(f)
for i, k in enumerate(testLabel.keys()):
#typePred[k,:] = testPred[i,:]
# some potential improvement: set a threshold for random walk number to block
# 'unconfident' data points
max = np.max(testPred[i, :])
if max > threshold[str(t)]:
typePred[k, :] = testPred[i, :]
# add meta-path probability to global probability
outPred += typePred * weight[str(t)]
with open(
'data/local/cotrain/' + scope_name + '/lb' +
str(lb).zfill(3) + '_pred_rd_' + str(rd).zfill(3) +
'_' + str(r).zfill(3), 'w') as f:
pk.dump(outPred, f)
return best_res