def test_census(self):
logger = get_logger('RL.Test.KmeansClustering.CENSUS')
census = Census()
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.candidate_links,
census.trainDataA, census.trainDataB)
logger.info("Features %s", str(features.describe()))
# Train K-Means Classifier
logrg = recordlinkage.KMeansClassifier(algorithm='full',
max_iter=1000,
random_state=42)
logrg.fit(features)
result = logrg.predict(features)
log_quality_results(logger, result, census.true_links,
len(census.candidate_links))
#Test the classifier
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.test_links, census.testDataA,
census.testDataB)
logger.info("Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, census.true_test_links,
len(census.test_links))
def test_febrl(self):
logger = get_logger('RL.Test.KmeansClustering.FEBRL')
febrl = FEBRL()
compare_cl = febrl.get_comparision_object()
features = compare_cl.compute(febrl.candidate_links, febrl.trainDataA,
febrl.trainDataB)
logger.info("Features %s", str(features.describe()))
# Train K-Means Classifier
logrg = recordlinkage.KMeansClassifier()
logrg.fit(features)
result = logrg.predict(features)
log_quality_results(logger, result, febrl.true_links,
len(febrl.candidate_links))
#Test the classifier
compare_cl = febrl.get_comparision_object()
features = compare_cl.compute(febrl.test_links, febrl.testDataA,
febrl.testDataB)
logger.info("Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, febrl.true_test_links,
len(febrl.test_links))
def test_cora(self):
logger = get_logger('RL.Test.KmeansClustering.CORA')
#Read Train data in dataset A & B
cora = Cora()
## Extarct Features
compare_cl = cora.get_comparision_object()
features = compare_cl.compute(cora.candidate_links, cora.trainDataA,
cora.trainDataB)
logger.info("Features %s", str(features.describe()))
# Train K-Means Classifier
logrg = recordlinkage.KMeansClassifier()
logrg.fit(features)
result = logrg.predict(features)
log_quality_results(logger, result, cora.true_links,
len(cora.candidate_links))
#Test the classifier
compare_cl = cora.get_comparision_object()
features = compare_cl.compute(cora.test_links, cora.testDataA,
cora.testDataB)
logger.info("Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, cora.true_test_links,
len(cora.test_links))
def _test_transh(self, dataset, params):
graph = Graph_ER(dataset)
model = dataset()
logger = get_logger('RL.Test.er.TransH.' + str(model))
transh = TransH(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = transh.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transh.get_ent_embeddings()
result_prob = []
for i in range(0, len(graph.entity_pairs)):
distance = abs(
spatial.distance.cosine(
ent_embeddings[graph.entity_pairs[i][0]],
ent_embeddings[graph.entity_pairs[i][1]]))
result_prob.append(
(graph.entity_pairs[i][0], graph.entity_pairs[i][1], distance))
#logger.info("i: %d, distance: %f true_pairs: %s", i, distance, graph.entity_pairs[i] in true_pairs)
#Write Embeddings to file
export_embeddings('er', str(model), 'TransH', graph.entity,
ent_embeddings)
export_result_prob(dataset, 'er', str(model), 'TransH', graph.entity,
result_prob, graph.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs)
try:
logger.info("MAX FSCORE: %f AT : %f", max_fscore,
optimal_threshold)
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
params['threshold'] = optimal_threshold
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
export_false_negatives(dataset, 'er', str(model), 'TransH',
graph.entity, result_prob, graph.true_pairs,
result, graph.entity)
export_false_positives(dataset, 'er', str(model), 'TransH',
graph.entity, result_prob, graph.true_pairs,
result, graph.entity)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
p_at_1 = ir_metrics.log_metrics(logger, params)
transh.close_tf_session()
return (max_fscore, p_at_1)
def _test_rl_transe(self, dataset, params):
#Load Graph Data
graph = Graph_ER(dataset)
model = dataset()
logger = get_logger('RL.Test.er.RLTransE.' + str(model))
transe = TransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transe.get_ent_embeddings()
result_prob = []
for (a, b) in graph.entity_pairs:
a_triples = [(h, t, r) for (h, t, r) in graph.triples if h == a]
b_triples = [(h, t, r) for (h, t, r) in graph.triples if h == b]
distance = abs(
spatial.distance.cosine(ent_embeddings[a], ent_embeddings[b]))
for (ah, at, ar) in a_triples:
bt = [t for (h, t, r) in b_triples if r == ar]
if len(bt):
distance = distance + abs(spatial.distance.cosine(\
ent_embeddings[at], ent_embeddings[bt[0]]))
result_prob.append((a, b, distance))
#logger.info("a: %d, b: %d distance: %f true_pairs: %s", a, b, distance, (a, b) in graph.true_pairs)
#Write Embeddings to file
export_embeddings('er', str(model), 'RLTransE', graph.entity,
ent_embeddings)
export_result_prob(dataset, 'er', str(model), 'RLTransE', graph.entity,
result_prob, graph.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs, max_threshold=3.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
precison_at_1 = ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
return (max_fscore, precison_at_1)
def _test_etranse(self, dataset, params):
model = dataset()
graph = Graph_ERER(dataset)
logger = get_logger("RL.Test.erer.ETransE." + str(model))
etranse = ETransE(graph,
dimension=params['dimension'],
batchSize=params['batchSize'],
learning_rate=params['learning_rate'],
margin=params['margin'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'],
regularizer_scale=params['regularizer_scale'],
alpha=params['alpha'],
beta=params['beta'])
etranse.train(max_epochs=params['max_epochs'])
ent_embeddings_a = etranse.get_ent_embeddings_A()
ent_embeddings_b = etranse.get_ent_embeddings_B()
result_prob = []
for i in range(0, len(graph.entity_pairs)):
distance = abs(
spatial.distance.cosine(
ent_embeddings_a[int(graph.entity_pairs[i][0])],
ent_embeddings_b[int(graph.entity_pairs[i][1])]))
result_prob.append(
(graph.entity_pairs[i][0], graph.entity_pairs[i][1], distance))
#logger.info("i: %d, distance: %f true_pairs: %s", i, distance, graph.entity_pairs[i] in true_pairs)
#Write Embeddings to file
export_embeddings('erer', str(model), 'ETransE', graph.entityA,
ent_embeddings_a)
export_embeddings('erer', str(model), 'ETransE', graph.entityB,
ent_embeddings_b)
export_result_prob(dataset, 'erer', str(model), 'ETransE',
graph.entityA, result_prob, graph.true_pairs,
graph.entityB)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
prec_at_1 = ir_metrics.log_metrics(logger, params)
etranse.close_tf_session()
return (max_fscore, prec_at_1)
def _test_erer(self, dataset, er_algo, params):
model = dataset()
graph = Graph_ERER(dataset)
graph_er = graph.get_er_model()
er_model = er_algo(graph_er,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = er_model.train(max_epochs=params['epochs'])
logger = get_logger('RL.Test.erer.ERER.' + str(model) + "." +
str(er_model))
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = er_model.get_ent_embeddings()
result_prob = []
for i in range(0, len(graph_er.entity_pairs)):
distance = abs(
spatial.distance.cosine(
ent_embeddings[graph_er.entity_pairs[i][0]],
ent_embeddings[graph_er.entity_pairs[i][1]]))
result_prob.append((graph_er.entity_pairs[i][0],
graph_er.entity_pairs[i][1], distance))
#logger.info("i: %d, distance: %f true_pairs: %s", i, distance, graph_er.entity_pairs[i] in graph_er.true_pairs)
#Write Embeddings to file
export_embeddings("erer", str(model), str(er_model), graph_er.entity,
ent_embeddings)
export_result_prob(dataset, 'erer', str(model), str(er_model),
graph_er.entity, result_prob, graph_er.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph_er.true_pairs)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph_er.true_pairs,
len(graph_er.entity_pairs), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob,
graph_er.true_pairs)
ir_metrics.log_metrics(logger)
er_model.close_tf_session()
return max_fscore
def _test_logistic_transh(self, dataset, params):
"""Note: Zero aligned pairs are returned, require fixation."""
model = dataset()
logger = get_logger('RL.Test.LogisticTransH.' + str(model))
entity, relation, triples, entity_pairs, true_pairs = model.get_er_model(
)
transh = TransH(entity,
relation,
triples,
entity_pairs,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
loss = transh.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transh.get_ent_embeddings()
ent_embeddings = [
np.array(ent_embeddings[i]) for i in range(ent_embeddings.shape[0])
]
trainDataA = pd.DataFrame(data=ent_embeddings)
trainDataB = pd.DataFrame(data=ent_embeddings)
compare_cl = recordlinkage.Compare()
for i in range(0, params['dimension']):
compare_cl.numeric(i, i, label=str(i), method='gauss')
candidate_links = pd.MultiIndex.from_tuples(entity_pairs)
features = compare_cl.compute(candidate_links, trainDataA, trainDataB)
logger.info("Features %s", str(features.describe()))
logrg = recordlinkage.LogisticRegressionClassifier()
logrg.fit(features, true_pairs)
result = logrg.predict(features)
log_quality_results(logger, result, true_pairs, len(entity_pairs))
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(entity_pairs[i][0], entity_pairs[i][1], prob[i])
for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob, true_pairs)
ir_metrics.log_metrics(logger, params)
def _test_seea(self, dataset, params):
model = dataset()
graph = Graph_EAR(dataset)
logger = get_logger('RL.Test.ear.SEEA.' + str(model))
seea = SEEA(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
batchSize=params['batchSize'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
#Begin SEEA iterations, passing true pairs only to debug the alignments.
results = seea.seea_iterate(beta=params['beta'],
max_iter=params['max_iter'],
max_epochs=params['max_epochs'])
try:
result_pairs = pd.MultiIndex.from_tuples(results)
fscore = log_quality_results(logger, result_pairs,
graph.true_pairs,
len(graph.entity_pairs), params)
except Exception as e:
logger.error(e)
logger.info("No Aligned pairs found.")
ent_embeddings = seea.get_ent_embeddings()
export_embeddings('ear', str(model), 'SEEA', graph.entity,
ent_embeddings)
result_prob = []
for (e1, e2) in graph.entity_pairs:
distance = abs(
spatial.distance.cosine(ent_embeddings[e1],
ent_embeddings[e2]))
result_prob.append((e1, e2, distance))
export_result_prob(dataset, 'ear', str(model), 'SEEA', graph.entity,
result_prob, graph.true_pairs)
try:
export_false_negatives(dataset, 'ear', str(model), 'SEEA',
graph.entity, result_prob, graph.true_pairs,
result_pairs, graph.entity)
export_false_positives(dataset, 'ear', str(model), 'SEEA',
graph.entity, result_prob, graph.true_pairs,
result_pairs, graph.entity)
except Exception as e:
logger.error(e)
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
prec_at_1 = ir_metrics.log_metrics(logger, params)
seea.close_tf_session()
return (fscore, prec_at_1)
def test_cora(self):
logger = get_logger('RL.Test.ECMClassifier.CORA')
#Read Train data in dataset A & B
cora = Cora()
## Extarct Features
compare_cl = cora.get_comparision_object()
features = compare_cl.compute(cora.candidate_links, cora.trainDataA,
cora.trainDataB)
logger.info("Train Features %s", str(features.describe()))
# Train ECM Classifier
logrg = recordlinkage.ECMClassifier()
logrg.fit(features)
result = logrg.predict(features)
log_quality_results(logger, result, cora.true_links,
len(cora.candidate_links))
#validate the classifier
compare_cl = cora.get_comparision_object()
features = compare_cl.compute(cora.val_links, cora.valDataA,
cora.valDataB)
logger.info("Validation Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, cora.true_val_links,
len(cora.val_links))
#Test the classifier
compare_cl = cora.get_comparision_object()
features = compare_cl.compute(cora.test_links, cora.testDataA,
cora.testDataB)
logger.info("Test Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, cora.true_test_links,
len(cora.test_links))
#Log IR Stats: MRR, MAP, MP@K
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(cora.test_links[i][0], cora.test_links[i][1], prob[i])
for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob,
cora.true_test_links)
ir_metrics.log_metrics(logger)
def test_febrl(self):
logger = get_logger('RL.Test.LogisticRegression.FEBRL')
febrl = FEBRL()
compare_cl = febrl.get_comparision_object()
features = compare_cl.compute(febrl.candidate_links, febrl.trainDataA,
febrl.trainDataB)
logger.info("Train Features %s", str(features.describe()))
# Train ECM Classifier
logrg = recordlinkage.LogisticRegressionClassifier()
logrg.fit(features, febrl.true_links)
result = logrg.predict(features)
log_quality_results(logger, result, febrl.true_links,
len(febrl.candidate_links))
#Validate the classifier
compare_cl = febrl.get_comparision_object()
features = compare_cl.compute(febrl.val_links, febrl.valDataA,
febrl.valDataB)
logger.info("Validation Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, febrl.true_val_links,
len(febrl.val_links))
#Test the classifier
compare_cl = febrl.get_comparision_object()
features = compare_cl.compute(febrl.test_links, febrl.testDataA,
febrl.testDataB)
logger.info("Test Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, febrl.true_test_links,
len(febrl.test_links))
#Log IR Stats: MRR, MAP, MP@K
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(febrl.test_links[i][0], febrl.test_links[i][1],
prob[i]) for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob,
febrl.true_test_links)
ir_metrics.log_metrics(logger)
def test_febrl(self, params=None):
if not params:
params = self.get_default_params()
#Load Graph Data
graph = Graph_ER(FEBRL)
model = FEBRL()
logger = get_logger('RL.Test.TransE.Household.' + str(model))
transe = TransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transe.get_ent_embeddings()
#Experimenting household matching
postcode_rel_id = graph.relation.index("postcode")
result_prob = []
for i in range(0, len(graph.entity_pairs)):
person_A = graph.entity_pairs[i][0]
person_B = graph.entity_pairs[i][1]
postcode_A = [
t for (h, t, r) in graph.triples
if h == person_A and r == postcode_rel_id
][0]
neighbours_A = [
h for (h, t, r) in graph.triples if t == postcode_A
]
#logger.info("FM A: %s", str([graph.entity[a] for a in neighbours_A]))
postcode_B = [
t for (h, t, r) in graph.triples
if h == person_B and r == postcode_rel_id
][0]
neighbours_B = [
h for (h, t, r) in graph.triples if t == postcode_B
]
#logger.info("FM B: %s", str([graph.entity[a] for a in neighbours_B]))
cost_matrix = np.zeros(shape=(len(neighbours_A),
len(neighbours_B)))
for i in range(len(neighbours_A)):
for j in range(len(neighbours_B)):
if neighbours_A[i] == neighbours_B[j]:
cost_matrix[i][j] = 100
else:
cost_matrix[i][j] = abs(
spatial.distance.cosine(
ent_embeddings[neighbours_A[i]],
ent_embeddings[neighbours_B[j]]))
#logger.info("Cost Matrix: %s", str(cost_matrix))
row_ind, col_ind = linear_sum_assignment(cost_matrix)
#logger.info("Cost of aligning = %f", cost_matrix[row_ind, col_ind].sum())
person_A_index = neighbours_A.index(person_A)
person_B_index = neighbours_B.index(person_B)
distance = cost_matrix[row_ind, col_ind].sum(
) + cost_matrix[person_A_index][person_B_index]
#import ipdb;ipdb.set_trace()
#if (person_A_index, person_B_index) not in (row_ind, col_ind):
# distance = distance + 1000
result_prob.append(
(graph.entity_pairs[i][0], graph.entity_pairs[i][1], distance))
export_embeddings('er', str(model), 'TransE.Household', graph.entity,
ent_embeddings)
export_result_prob(FEBRL, 'er', str(model), 'TransE.Household',
graph.entity, result_prob, graph.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
return max_fscore
def test_census(self, params=None):
if not params:
params = self.get_default_params()
#Load Graph Data
graph = Graph_ER(Census)
model = Census()
logger = get_logger('RL.Test.TransE.Household.' + str(model))
transe = TransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transe.get_ent_embeddings()
#Experimenting household matching
result_prob = []
for ep_index in range(0, len(graph.entity_pairs)):
#logger.info("Computing cost for: %s", str([graph.entity[e] for e in graph.entity_pairs[ep_index]]))
household_A = [
t for (h, t, r) in graph.triples
if h == graph.entity_pairs[ep_index][0] and r > 6
][0]
family_members_A = [
h for (h, t, r) in graph.triples if t == household_A
]
#logger.info("FM A: %s", str([graph.entity[a] for a in family_members_A]))
household_B = [
t for (h, t, r) in graph.triples
if h == graph.entity_pairs[ep_index][1] and r > 6
][0]
family_members_B = [
h for (h, t, r) in graph.triples if t == household_B
]
#logger.info("FM B: %s", str([graph.entity[a] for a in family_members_B]))
cost_matrix = np.zeros(shape=(len(family_members_A),
len(family_members_B)))
for i in range(len(family_members_A)):
for j in range(len(family_members_B)):
#if family_members_A[i] == family_members_B[j]:
# cost_matrix[i][j] = 100
#else:
cost_matrix[i][j] = abs(
spatial.distance.cosine(
ent_embeddings[family_members_A[i]],
ent_embeddings[family_members_B[j]]))
#logger.info("Cost Matrix: %s", str(cost_matrix))
row_ind, col_ind = linear_sum_assignment(cost_matrix)
#logger.info("Cost of aligning = %f", cost_matrix[row_ind, col_ind].sum())
#logger.info("Rows selected %s, Col selected: %s", str(row_ind), str(col_ind))
eA_index = family_members_A.index(graph.entity_pairs[ep_index][0])
eB_index = family_members_B.index(graph.entity_pairs[ep_index][1])
#logger.info("A index: %d, B index: %d", eA_index, eB_index)
rowA = np.where(row_ind == eA_index)[0]
if len(rowA) and col_ind[rowA[0]] == eB_index:
#logger.info("Pair in min. cost matrix")
distance = cost_matrix[row_ind, col_ind].sum()
else:
distance = cost_matrix[row_ind, col_ind].sum() + abs(
spatial.distance.cosine(
ent_embeddings[graph.entity_pairs[ep_index][0]],
ent_embeddings[graph.entity_pairs[ep_index][1]]))
result_prob.append(
(graph.entity_pairs[i][0], graph.entity_pairs[i][1], distance))
if ep_index % 1000 == 0:
logger.info("i: %d, distance: %f true_pairs: %s", ep_index,
distance, graph.entity_pairs[ep_index]
in graph.true_pairs)
#if graph.entity_pairs[ep_index] in graph.true_pairs:
# import ipdb;ipdb.set_trace()
#Normalize distance
max_distance = 10
#for r in result_prob:
# if r[2] > max_distance:
# max_distance = r[2]
result_prob = [(r[0], r[1], (r[2] / max_distance))
for r in result_prob]
#logger.info("Max distance: %f", max_distance)
for r in result_prob[:100]:
logger.info("distance: %f true_pairs: %s", r[2], (r[0], r[1])
in graph.true_pairs)
export_embeddings('er', str(model), 'TransE.Household', graph.entity,
ent_embeddings)
export_result_prob(Census, 'er', str(model), 'TransE.Household',
graph.entity, result_prob, graph.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
return max_fscore
def test_census_new(self):
c = Census()
graph = Graph_VEG(Census)
logger = get_logger("RL.Test.LogisticRLTransE.Census")
logger.info("values for name : %s",
str(graph.relation_value_map[graph.relation[1]][:10]))
logger.info("relation: %s", str(graph.relation))
logger.info("train_triples: %s", str(graph.train_triples[:10]))
logger.info("set train_triples size %d", len(set(graph.train_triples)))
params = self.get_default_params()
transe = RLTransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss, val_loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f val_loss: %f", loss,
val_loss)
value_embeddings = transe.get_val_embeddings()
relation_embeddings = transe.get_rel_embeddings()
#Map of feilds in census dataFrame to VEG relations.
field_relation_map = {
c.field_map[CensusFields.FIRST_NAME]: "name",
c.field_map[CensusFields.SURNAME_1]: "surname",
c.field_map[CensusFields.SURNAME_2]: "surname2",
c.field_map[CensusFields.YOB]: "yob",
c.field_map[CensusFields.CIVIL_STATUS]: "civil",
c.field_map[CensusFields.OCCUPATION]: "occupation",
c.field_map[CensusFields.RELATION]: "relation"
}
missing_values = []
train_features = [] #Size samples*(dimension*rel_count)
test_features = []
for (candidate_links, dataA, dataB, features) in \
[(c.candidate_links, c.trainDataA, c.trainDataB, train_features),
(c.test_links, c.testDataA, c.testDataB, test_features)]:
for (a, b) in candidate_links:
row_a = dataA.loc[a]
row_b = dataB.loc[b]
distance = []
for f in field_relation_map:
val_a = row_a[f]
val_b = row_b[f]
if val_a != val_b:
rel = field_relation_map[f]
try:
val_index_a = graph.relation_value_map[rel].index(
val_a)
except ValueError:
missing_values.append(val_a)
distance.extend([1.0] * params['dimension'])
continue
try:
val_index_b = graph.relation_value_map[rel].index(
val_b)
except ValueError:
missing_values.append(val_b)
distance.extend([1.0] * params['dimension'])
continue
rel_index = graph.relation.index(field_relation_map[f])
distance.extend(value_embeddings[rel][val_index_a] + \
relation_embeddings[rel_index] - value_embeddings[rel][val_index_b])
features.append(pd.Series(distance).rename((a, b)))
#logger.info("a: %d, b: %d distance: %f true_pairs: %s", a, b, distance, (a,b) in c.true_test_links)
logger.info("No. of missing values: %d", len(missing_values))
logger.info("Unique No. of missing values: %d",
len(set(missing_values)))
train_features = pd.DataFrame(data=train_features).fillna(1)
test_features = pd.DataFrame(data=test_features).fillna(1)
logger.info("Shape of Train features: %s", str(train_features.shape))
logger.info("Shape of Test features: %s", str(test_features.shape))
#Train Logistic Regression Model
logrg = recordlinkage.LogisticRegressionClassifier()
logrg.fit(train_features, c.true_links)
result = logrg.predict(train_features)
result = pd.MultiIndex.from_tuples(result.to_series())
log_quality_results(logger, result, c.true_links,
len(c.candidate_links), params)
#Test Classifier
result = logrg.predict(test_features)
result = pd.MultiIndex.from_tuples(result.to_series())
log_quality_results(logger, result, c.true_test_links,
len(c.test_links), params)
"""
Todo: Export Embeddings and probabilities.
try:
entities = ["value\trelation"]
for r in graph.relation_value_map:
for v in graph.relation_value_map[r]:
entities.append("\t".join([v,r]))
embeddings = []
for rel in value_embeddings:
val_count = len(graph.relation_value_map[rel])
embeddings.extend(value_embeddings[rel][:val_count])
#Write Embeddings to file
export_embeddings('veg', str(c), 'LogisticRLTransE', entities, embeddings)
except Exception as e:
logger.error("Failed to export embeddings")
logger.error(e)
export_result_prob(Census, 'veg', str(c), 'RLTransE', graph.values, result_prob, c.true_test_links)
"""
prob_series = logrg.prob(test_features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(c.test_links[i][0], c.test_links[i][1], prob[i])
for i in range(0, len(prob))]
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob,
c.true_test_links)
precison_at_1 = ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
def _test_rl_transe(self, model, field_relation_map, params):
dataset = model()
graph = Graph_VEG(model)
logger = get_logger("RL.Test.RLTransE." + str(dataset))
logger.info("values for name : %s",
str(graph.relation_value_map[graph.relation[1]][:10]))
logger.info("relation: %s", str(graph.relation))
logger.info("train_triples: %s", str(graph.train_triples[:10]))
logger.info("set train_triples size %d", len(set(graph.train_triples)))
transe = RLTransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss, val_loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f val_loss: %f", loss,
val_loss)
value_embeddings = transe.get_val_embeddings()
relation_embeddings = transe.get_rel_embeddings()
result_prob = []
distance_distribution = []
missing_values = []
for (a, b) in dataset.test_links:
row_a = dataset.testDataA.loc[a]
row_b = dataset.testDataB.loc[b]
distance = 0
dd = []
for f in field_relation_map:
val_a = row_a[f]
val_b = row_b[f]
if val_a == val_b:
dd.append(0)
else:
rel = field_relation_map[f]
try:
val_index_a = graph.relation_value_map[rel].index(
val_a)
except ValueError:
missing_values.append(val_a)
distance = distance + 1
dd.append(1)
continue
try:
val_index_b = graph.relation_value_map[rel].index(
val_b)
except ValueError:
missing_values.append(val_b)
distance = distance + 1
dd.append(1)
continue
rel_index = graph.relation.index(field_relation_map[f])
cur_distance = abs(
spatial.distance.cosine(
value_embeddings[rel][val_index_a] +
relation_embeddings[rel_index],
value_embeddings[rel][val_index_b]))
distance = distance + cur_distance
dd.append(cur_distance)
result_prob.append((a, b, distance))
distance_distribution.append((a, b, dd, distance))
#logger.info("a: %d, b: %d distance: %f true_pairs: %s", a, b, distance, (a,b) in dataset.true_test_links)
logger.info("No. of missing values: %d", len(missing_values))
logger.info("Unique No. of missing values: %d",
len(set(missing_values)))
try:
entities = ["value\trelation"]
for r in graph.relation_value_map:
for v in graph.relation_value_map[r]:
entities.append("\t".join([v, r]))
embeddings = []
for rel in value_embeddings:
val_count = len(graph.relation_value_map[rel])
embeddings.extend(value_embeddings[rel][:val_count])
#Write Embeddings to file
export_embeddings('veg', str(dataset), 'RLTransE_val', entities,
embeddings)
export_embeddings('veg', str(dataset), 'RLTransE_rel',
graph.relation, relation_embeddings)
except Exception as e:
logger.error("Failed to export embeddings")
logger.error(e)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=3.0, step=0.02)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob,
dataset.true_test_links)
precison_at_1 = ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
#Export False Positives and result porobabilities
get_entity_name = lambda d, i: "_".join([
str(d.iloc[i][dataset.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.iloc[i][dataset.field_map[CensusFields.DNI]])
])
get_entity_name_loc = lambda d, i: "_".join([
str(d.loc[i][dataset.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.loc[i][dataset.field_map[CensusFields.DNI]])
])
entitiesA = [
get_entity_name(dataset.testDataA, i)
for i in range(int(dataset.testDataA.shape[0]))
]
entitiesB = [
get_entity_name(dataset.testDataB, i)
for i in range(int(dataset.testDataB.shape[0]))
]
result_prob = [
(entitiesA.index(get_entity_name_loc(dataset.testDataA, int(a))),
entitiesB.index(get_entity_name_loc(dataset.testDataB,
int(b))), p)
for (a, b, p) in result_prob
]
true_links = [
(entitiesA.index(get_entity_name_loc(dataset.testDataA, int(a))),
entitiesB.index(get_entity_name_loc(dataset.testDataB, int(b))))
for (a, b) in dataset.true_test_links
]
export_result_prob(Census, 'veg', 'census', 'rltranse', entitiesA,
result_prob, true_links, entitiesB)
distance_distribution = [
(entitiesA.index(get_entity_name_loc(dataset.testDataA, int(a))),
entitiesB.index(get_entity_name_loc(dataset.testDataB, int(b))),
[str("%.2f" % (float(w))) for w in dd], 1 - d)
for (e1, e2, dd, d) in distance_distribution if (e1, e2) in result
]
export_human_readable_results(Census, 'veg', 'census', 'rltranse',
entitiesA, distance_distribution,
entitiesB)
result = [
(entitiesA.index(get_entity_name_loc(dataset.testDataA, int(a))),
entitiesB.index(get_entity_name_loc(dataset.testDataB, int(b))))
for (a, b) in result
]
export_false_negatives(Census, 'veg', 'census', 'rltranse', entitiesA,
result_prob, true_links, result, entitiesB)
export_false_positives(Census, 'veg', 'census', 'rltranse', entitiesA,
result_prob, true_links, result, entitiesB)
return (max_fscore, precison_at_1)
def test_ecm(self):
logger = get_logger('RL.Test.ECMClassifier.Census')
census = Census()
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.candidate_links,
census.trainDataA, census.trainDataB)
logger.info("Train Features %s", str(features.describe()))
# Train ECM Classifier
logrg = recordlinkage.ECMClassifier()
logrg.fit(features)
result = logrg.predict(features)
log_quality_results(logger, result, census.true_links,
len(census.candidate_links))
#Validate the classifier
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.val_links, census.valDataA,
census.valDataB)
logger.info("Validation Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, census.true_val_links,
len(census.val_links))
#Test the classifier
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.test_links, census.testDataA,
census.testDataB)
logger.info("Test Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, census.true_test_links,
len(census.test_links))
logger.info("ECM weights: %s", str(logrg.weights))
#Log IR Stats: MRR, MAP, MP@K
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(census.test_links[i][0], census.test_links[i][1],
prob[i]) for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob,
census.true_test_links)
ir_metrics.log_metrics(logger)
#Export False Positives and result porobabilities
result_feature_mapping = [
(e1, e2, [str(v) for v in features.loc[(e1, e2)].values], d)
for (e1, e2, d) in result_prob if (e1, e2) in result
]
get_entity_name = lambda c, d, i: "_".join([
str(d.iloc[i][c.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.iloc[i][c.field_map[CensusFields.DNI]])
])
get_entity_name_loc = lambda c, d, i: "_".join([
str(d.loc[i][c.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.loc[i][c.field_map[CensusFields.DNI]])
])
start_time = timeit.default_timer()
entitiesA = [
get_entity_name(census, census.testDataA, i)
for i in range(int(census.testDataA.shape[0]))
]
entitiesB = [
get_entity_name(census, census.testDataB, i)
for i in range(int(census.testDataB.shape[0]))
]
logger.info("Entities built in %s",
str(timeit.default_timer() - start_time))
start_time = timeit.default_timer()
result_prob = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB,
int(b))), p)
for (a, b, p) in result_prob]
logger.info("Result prob in %s",
str(timeit.default_timer() - start_time))
start_time = timeit.default_timer()
true_links = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB,
int(b))))
for (a, b) in census.true_test_links]
logger.info("true_links in %s",
str(timeit.default_timer() - start_time))
start_time = timeit.default_timer()
export_result_prob(Census, 'ECM', 'census', 'ecm', entitiesA,
result_prob, true_links, entitiesB)
logger.info("Result prob EXPORTED in %s",
str(timeit.default_timer() - start_time))
start_time = timeit.default_timer()
result = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB, int(b))))
for (a, b) in result]
export_false_negatives(Census, 'ECM', 'census', 'ecm', entitiesA,
result_prob, true_links, result, entitiesB)
export_false_positives(Census, 'ECM', 'census', 'ecm', entitiesA,
result_prob, true_links, result, entitiesB)
logger.info("FP & FN EXPORTED in %s",
str(timeit.default_timer() - start_time))
result_feature_mapping = [
(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB, int(b))), w, p)
for (a, b, w, p) in result_feature_mapping
]
export_human_readable_results(Census, 'ECM', 'census', 'ecm',
entitiesA, result_feature_mapping,
entitiesB)
logger.info("Exported Human Readable Results")
def _test_logistic_transh_erer(self, dataset, params):
model = dataset()
logger = get_logger('RL.Test.erer.LogisticTransH.ERER.' + str(model))
entA, entB, relA, relB, triA, triB, entity_pairs, prior_pairs, true_pairs = model.get_erer_model(
)
self.assertTrue(all([(tp in entity_pairs) for tp in true_pairs]))
#Generate embeddings for datasetA
transh = TransH(entA,
relA,
triA,
prior_pairs,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
loss = transh.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddingsA = transh.get_ent_embeddings()
transh.close_tf_session()
del transh
#Generate embeddings for datasetB
transh = TransH(entB,
relB,
triB,
entity_pairs,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
loss = transh.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddingsB = transh.get_ent_embeddings()
transh.close_tf_session()
ent_embeddingsA = [
np.array(ent_embeddingsA[i])
for i in range(ent_embeddingsA.shape[0])
]
ent_embeddingsB = [
np.array(ent_embeddingsB[i])
for i in range(ent_embeddingsB.shape[0])
]
trainDataA = pd.DataFrame(data=ent_embeddingsA)
trainDataB = pd.DataFrame(data=ent_embeddingsB)
#Define comparision Class
compare_cl = recordlinkage.Compare()
for i in range(0, params['dimension']):
compare_cl.numeric(i, i, label=str(i)) #method='exp')
#sample negative pairs
train_pairs = []
tuple_pp = set(map(tuple, prior_pairs))
logger.info("Number of prior_pairs: %d", len(prior_pairs))
for e1, e2 in prior_pairs:
train_pairs.append((e1, e2))
while True:
neg_e2 = random.choice(xrange(0, len(entB)))
if neg_e2 == e2 or (e1, neg_e2) in tuple_pp:
continue
else:
train_pairs.append((e1, neg_e2))
break
logger.info("Number of Train Pairs: %d", len(train_pairs))
candidate_links = pd.MultiIndex.from_tuples(train_pairs)
features = compare_cl.compute(candidate_links, trainDataA, trainDataB)
logger.info("Train Features %s", str(features.describe()))
#Train Logistic Regression Model
logrg = recordlinkage.LogisticRegressionClassifier()
candidate_links = pd.MultiIndex.from_tuples(prior_pairs)
logrg.fit(features, candidate_links)
#Test Classifier
compare_cl = recordlinkage.Compare()
for i in range(0, params['dimension']):
compare_cl.numeric(i, i, label=str(i))
candidate_links = pd.MultiIndex.from_tuples(entity_pairs)
features = compare_cl.compute(candidate_links, trainDataA, trainDataB)
logger.info("Test Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, true_pairs, len(entity_pairs))
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(entity_pairs[i][0], entity_pairs[i][1], prob[i])
for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob, true_pairs)
ir_metrics.log_metrics(logger, params, params)
#Export results
export_embeddings('erer', str(model), 'LogTransH', entA,
ent_embeddingsA)
export_embeddings('erer', str(model), 'LogTransH', entB,
ent_embeddingsB)
export_result_prob(dataset, 'erer', str(model), 'LogTransH', entA,
result_prob, true_pairs, entB)
def test_logistic(self):
logger = get_logger('RL.Test.LogisticRegression.Census')
census = Census()
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.candidate_links,
census.trainDataA, census.trainDataB)
logger.info("Train Features %s", str(features.describe()))
# Train ECM Classifier
logrg = recordlinkage.LogisticRegressionClassifier()
logrg.fit(features, census.true_links)
result = logrg.predict(features)
log_quality_results(logger, result, census.true_links,
len(census.candidate_links))
#Validate the classifier
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.val_links, census.valDataA,
census.valDataB)
logger.info("Validation Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, census.true_val_links,
len(census.val_links))
#Test the classifier
compare_cl = census.get_comparision_object()
features = compare_cl.compute(census.test_links, census.testDataA,
census.testDataB)
logger.info("Test Features %s", str(features.describe()))
result = logrg.predict(features)
log_quality_results(logger, result, census.true_test_links,
len(census.test_links))
logger.info("logrg coefficients: %s", str(logrg.coefficients))
#Log IR Stats: MRR, MAP, MP@K
prob_series = logrg.prob(features)
prob = [(1 - p) for p in prob_series.tolist()]
result_prob = [(census.test_links[i][0], census.test_links[i][1],
prob[i]) for i in range(0, len(prob))]
ir_metrics = InformationRetrievalMetrics(result_prob,
census.true_test_links)
ir_metrics.log_metrics(logger)
#Export False Positives and result porobabilities
result_feature_mapping = [
(e1, e2, [str(v) for v in features.loc[(e1, e2)].values], d)
for (e1, e2, d) in result_prob if (e1, e2) in result
]
get_entity_name = lambda c, d, i: "_".join([
str(d.iloc[i][c.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.iloc[i][c.field_map[CensusFields.DNI]])
])
get_entity_name_loc = lambda c, d, i: "_".join([
str(d.loc[i][c.field_map[CensusFields.ID_INDIVIDUAL]]),
str(d.loc[i][c.field_map[CensusFields.DNI]])
])
entitiesA = [
get_entity_name(census, census.testDataA, i)
for i in range(int(census.testDataA.shape[0]))
]
entitiesB = [
get_entity_name(census, census.testDataB, i)
for i in range(int(census.testDataB.shape[0]))
]
result_prob = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB,
int(b))), p)
for (a, b, p) in result_prob]
true_links = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB,
int(b))))
for (a, b) in census.true_test_links]
export_result_prob(Census, 'LogisticRegression', 'census', 'logistic',
entitiesA, result_prob, true_links, entitiesB)
result = [(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB, int(b))))
for (a, b) in result]
export_false_negatives(Census, 'LogisticRegression', 'census',
'logistic', entitiesA, result_prob, true_links,
result, entitiesB)
export_false_positives(Census, 'LogisticRegression', 'census',
'logistic', entitiesA, result_prob, true_links,
result, entitiesB)
weights = logrg.coefficients
result = [
(e1, e2,
[str("%.2f" % (float(d * w) / sum(weights))) for w in weights], d)
for (e1, e2, d) in result_prob if (e1, e2) in result
]
result_feature_mapping = [
(entitiesA.index(
get_entity_name_loc(census, census.testDataA, int(a))),
entitiesB.index(
get_entity_name_loc(census, census.testDataB, int(b))), w, p)
for (a, b, w, p) in result_feature_mapping
]
export_human_readable_results(Census, 'LogisticRegression', 'census',
'logistic', entitiesA,
result_feature_mapping, entitiesB)
def test_cora(self, params=None):
if not params:
params = self.get_default_params()
#Load Graph Data
graph = Graph_ER(Cora)
model = Cora()
logger = get_logger('RL.Test.TransE.Household.' + str(model))
transe = TransE(graph,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'],
neg_rate=params['neg_rate'],
neg_rel_rate=params['neg_rel_rate'])
loss = transe.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f", loss)
ent_embeddings = transe.get_ent_embeddings()
#Experimenting household matching
auth_rel_index = graph.relation.index('author')
result_prob = []
for ep_index in range(0, len(graph.entity_pairs)):
authors_A = [
t for (h, t, r) in graph.triples
if h == graph.entity_pairs[ep_index][0] and r == auth_rel_index
]
#logger.info("AUHTORS A: %s", str([graph.entity[a] for a in authors_A]))
authors_B = [
t for (h, t, r) in graph.triples
if h == graph.entity_pairs[ep_index][1] and r == auth_rel_index
]
#logger.info("AUHTORS B: %s", str([graph.entity[a] for a in authors_B]))
cost_matrix = np.zeros(shape=(len(authors_A), len(authors_B)))
for i in range(len(authors_A)):
for j in range(len(authors_B)):
#if authors_A[i] == authors_B[j]:
# cost_matrix[i][j] = 100
#else:
cost_matrix[i][j] = abs(
spatial.distance.cosine(ent_embeddings[authors_A[i]],
ent_embeddings[authors_B[j]]))
#logger.info("Cost Matrix: %s", str(cost_matrix))
row_ind, col_ind = linear_sum_assignment(cost_matrix)
#logger.info("Cost of aligning = %f", cost_matrix[row_ind, col_ind].sum())
distance = cost_matrix[row_ind, col_ind].sum() + abs(
spatial.distance.cosine(
ent_embeddings[graph.entity_pairs[ep_index][0]],
ent_embeddings[graph.entity_pairs[ep_index][1]]))
result_prob.append((graph.entity_pairs[ep_index][0],
graph.entity_pairs[ep_index][1], distance))
if distance <= 0.05:
logger.info("i: %d, distance: %f true_pairs: %s", ep_index,
distance, graph.entity_pairs[ep_index]
in graph.true_pairs)
export_embeddings('er', str(model), 'TransE.Household', graph.entity,
ent_embeddings)
export_result_prob(Cora, 'er', str(model), 'TransE.Household',
graph.entity, result_prob, graph.true_pairs)
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, graph.true_pairs)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, graph.true_pairs,
len(graph.entity_pairs), params)
except:
logger.info("Zero Reults")
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob, graph.true_pairs)
ir_metrics.log_metrics(logger, params)
transe.close_tf_session()
return max_fscore
def test_veer(self):
logger = get_logger('RL.Test.VEER.Census')
dataset = Census()
#Columns of interest for Sant Feliu town
columns = [
'Noms_harmo', 'cognom_1', 'cohort', 'estat_civil', 'parentesc_har',
'ocupacio_hisco'
]
params = {
'learning_rate': 0.1,
'margin': 0.1,
'dimension': 32,
'epochs': 50,
'regularizer_scale': 0.1,
'batchSize': 512
}
veer = VEER(Census,
columns,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
#Train Model
loss, val_loss = veer.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f, val_loss:%f", loss,
val_loss)
#Test Model
result_prob, accuracy = veer.test()
logger.info("Predict count: %d", len(result_prob))
logger.info(
"Sample Prob: %s",
str([(c, (a, b) in dataset.true_test_links)
for (a, b, c) in result_prob[:20]]))
logger.info("Column Weights: %s", str(veer.get_col_weights()))
logger.info("Accuracy: %s", str(accuracy))
logger.info("Sample embeddings: %s", str(veer.get_val_embeddings()[0]))
#Compute Performance measures
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=2.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob,
dataset.true_test_links)
precison_at_1 = ir_metrics.log_metrics(logger, params)
#Export embeddings
embeddings = veer.get_val_embeddings()
export_embeddings('veg', 'census', 'veer', veer.values, embeddings)
#Write Result Prob to file
result_feature_mapping = [(e1, e2, [
str(
abs(
spatial.distance.cosine(
embeddings[veer.values.index(
veer._clean(dataset.testDataA.loc[e1][c]))],
embeddings[veer.values.index(
veer._clean(dataset.testDataB.loc[e2][c]))])))
for c in columns
], d) for (e1, e2, d) in result_prob if (e1, e2) in result]
entitiesA = dataset.get_entity_names(dataset.testDataA)
entitiesB = dataset.get_entity_names(dataset.testDataB)
index_dictA = {
str(dataset.testDataA.iloc[i]._name): i
for i in range(dataset.testDataA.shape[0])
}
index_dictB = {
str(dataset.testDataB.iloc[i]._name): i
for i in range(dataset.testDataB.shape[0])
}
result_prob = [(index_dictA[str(a)], index_dictB[str(b)], p)
for (a, b, p) in result_prob]
export_result_prob(dataset, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, dataset.true_test_links, entitiesB)
export_false_negatives(Census, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, dataset.true_test_links, result,
entitiesB)
export_false_positives(Census, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, dataset.true_test_links, result,
entitiesB)
result_feature_mapping = [(index_dictA[str(a)], index_dictB[str(b)], w,
p)
for (a, b, w, p) in result_feature_mapping]
export_human_readable_results(Census, 'veg', str(dataset), 'VEER',
entitiesA, result_feature_mapping,
entitiesB)
veer.close_tf_session()
def _test_veer(self, model, columns, params):
#Load Graph Data
dataset = model()
logger = get_logger('RL.Test.VEER.' + str(dataset))
veer = VEER(model,
columns,
dimension=params['dimension'],
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
#Train Model
loss, val_loss = veer.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f, val_loss:%f", loss,
val_loss)
#Test Model
result_prob, accuracy = veer.test()
logger.info("Predict count: %d", len(result_prob))
logger.info(
"Sample Prob: %s",
str([(c, (a, b) in dataset.true_test_links)
for (a, b, c) in result_prob[:20]]))
logger.info("Column Weights: %s", str(veer.get_col_weights()))
logger.info("Accuracy: %s", str(accuracy))
logger.info("Sample embeddings: %s", str(veer.get_val_embeddings()[0]))
#Compute Performance measures
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=2.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Log MAP, MRR and Hits@K
ir_metrics = InformationRetrievalMetrics(result_prob,
dataset.true_test_links)
precison_at_1 = ir_metrics.log_metrics(logger, params)
#Write Result Prob to file
entitiesA = dataset.get_entity_names(dataset.testDataA)
entitiesB = dataset.get_entity_names(dataset.testDataB)
index_dictA = {
str(dataset.testDataA.iloc[i]._name): i
for i in range(dataset.testDataA.shape[0])
}
index_dictB = {
str(dataset.testDataB.iloc[i]._name): i
for i in range(dataset.testDataB.shape[0])
}
result_prob = [(index_dictA[str(a)], index_dictB[str(b)], p)
for (a, b, p) in result_prob]
true_links = [(index_dictA[str(a)], index_dictB[str(b)])
for (a, b) in dataset.true_test_links]
export_result_prob(dataset, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, true_links, entitiesB)
result = [(index_dictA[str(a)], index_dictB[str(b)])
for (a, b) in result]
export_false_negatives(model, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, true_links, result, entitiesB)
export_false_positives(model, 'veg', str(dataset), 'VEER', entitiesA,
result_prob, true_links, result, entitiesB)
veer.close_tf_session()
return (max_fscore, precison_at_1)
def _test_werl(self, model, columns, params):
#Load Graph Data
dataset = model()
logger = get_logger('RL.Test.WERL.' + str(dataset))
ea_params = self.get_optimal_ea_params(model, params['ea_method'])
if params['ea_method'] in [TransE, TransH]:
#ER methods
graph = Graph_ER(model)
#Train TransE embedding vectors
transe = params['ea_method'](
graph,
dimension=ea_params['dimension'],
learning_rate=ea_params['learning_rate'],
margin=ea_params['margin'],
regularizer_scale=ea_params['regularizer_scale'],
batchSize=ea_params['batchSize'],
neg_rate=ea_params['neg_rate'],
neg_rel_rate=ea_params['neg_rel_rate'])
try:
#raise Exception("Reset")
transe.restore_model(
self._get_tf_model_filename(dataset, transe))
except Exception as e:
logger.error(e)
loss = transe.train(max_epochs=ea_params['epochs'])
logger.info("Training Complete with loss: %f", loss)
transe.save_model(self._get_tf_model_filename(dataset, transe))
ent_embeddings = transe.get_ent_embeddings()
rel_embeddings = None
entity = graph.entity
transe.close_tf_session()
elif params['ea_method'] in [RLTransE]:
#VEG methods
graph = Graph_VEG(model)
#Train TransE embedding vectors
rltranse = params['ea_method'](
graph,
dimension=ea_params['dimension'],
learning_rate=ea_params['learning_rate'],
margin=ea_params['margin'],
regularizer_scale=ea_params['regularizer_scale'],
batchSize=ea_params['batchSize'],
neg_rate=ea_params['neg_rate'],
neg_rel_rate=ea_params['neg_rel_rate'])
try:
#raise Exception("Reset")
rltranse.restore_model(
self._get_tf_model_filename(dataset, rltranse))
except Exception as e:
logger.error(e)
loss, val_loss = rltranse.train(max_epochs=ea_params['epochs'])
logger.info("Training Complete with loss: %f", loss)
rltranse.save_model(
self._get_tf_model_filename(dataset, rltranse))
val_embeddings = rltranse.get_val_embeddings()
rel_embeddings = rltranse.get_rel_embeddings()
if model == Census:
#hack: census veg graph has 8 relations. we need only 6
#removing same_as and surname2 embedding.
rel_embeddings = np.append(rel_embeddings[1:3],
rel_embeddings[4:],
axis=0)
ent_embeddings = []
entity = []
for rel in val_embeddings:
val_count = len(graph.relation_value_map[rel])
entity.extend(graph.relation_value_map[rel])
ent_embeddings.extend(val_embeddings[rel][:val_count])
assert len(ent_embeddings) == len(entity)
rltranse.close_tf_session()
elif params['ea_method'] in [VEER]:
veer = VEER(model,
columns,
dimension=ea_params['dimension'],
learning_rate=ea_params['learning_rate'],
margin=ea_params['margin'],
regularizer_scale=ea_params['regularizer_scale'],
batchSize=ea_params['batchSize'])
try:
veer.restore_model(self._get_tf_model_filename(dataset, veer))
except Exception as e:
logger.error(e)
#Train Model
loss, val_loss = veer.train(max_epochs=ea_params['epochs'])
logger.info("Training Complete with loss: %f, val_loss:%f",
loss, val_loss)
veer.save_model(self._get_tf_model_filename(dataset, veer))
ent_embeddings = veer.get_val_embeddings()
rel_embeddings = None
entity = veer.get_values()
veer.close_tf_session()
else:
raise Exception("Unknown Entity Alignment method")
#Train WERL weights
werl = WERL(model,
columns,
entity,
ent_embeddings,
rel_embeddings,
learning_rate=params['learning_rate'],
margin=params['margin'],
regularizer_scale=params['regularizer_scale'],
batchSize=params['batchSize'])
loss, val_loss = werl.train(max_epochs=params['epochs'])
logger.info("Training Complete with loss: %f, val_loss:%f", loss,
val_loss)
#Test Model
result_prob, accuracy = werl.test()
logger.info("Predict count: %d", len(result_prob))
logger.info(
"Sample Prob: %s",
str([(c, (a, b) in dataset.true_test_links)
for (a, b, c) in result_prob[:20]]))
logger.info("Column Weights: %s", str(werl.get_col_weights()))
logger.info("Accuracy: %s", str(accuracy))
#Compute Performance measures
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=2.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Test Model
logger = get_logger('RL.Test.MERL.' + str(dataset))
result_prob, accuracy = werl.test_merl()
logger.info("Predict count: %d", len(result_prob))
logger.info(
"Sample Prob: %s",
str([(c, (a, b) in dataset.true_test_links)
for (a, b, c) in result_prob[:20]]))
logger.info("Column Weights: %s", str(werl.get_col_weights()))
logger.info("Accuracy: %s", str(accuracy))
#Compute Performance measures
optimal_threshold, max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=2.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Log MAP, MRR and Hits@K
#ir_metrics = InformationRetrievalMetrics(result_prob, dataset.true_test_links)
precison_at_1 = None #ir_metrics.log_metrics(logger, params)
#Test Without Weights = Mean Emebedding for Record Linkage
logger = get_logger('RL.Test.NoWT.' + str(dataset))
result_prob, accuracy = werl.test_without_weight()
logger.info("Predict count: %d", len(result_prob))
logger.info(
"Sample Prob: %s",
str([(c, (a, b) in dataset.true_test_links)
for (a, b, c) in result_prob[:20]]))
logger.info("Column Weights: %s", str(werl.get_col_weights()))
logger.info("Accuracy: %s", str(accuracy))
#Compute Performance measures
optimal_threshold, nowt_max_fscore = get_optimal_threshold(
result_prob, dataset.true_test_links, max_threshold=2.0)
try:
params['threshold'] = optimal_threshold
result = pd.MultiIndex.from_tuples([(e1, e2)
for (e1, e2, d) in result_prob
if d <= optimal_threshold])
log_quality_results(logger, result, dataset.true_test_links,
len(dataset.test_links), params)
except Exception as e:
logger.info("Zero Reults")
logger.error(e)
#Log MAP, MRR and Hits@K
#ir_metrics = InformationRetrievalMetrics(result_prob, dataset.true_test_links)
#nowt_precison_at_1 = ir_metrics.log_metrics(logger, params)
werl.close_tf_session()
return (max_fscore, precison_at_1)
