def parse_ms(s):
fn_data = Name_functions.DS_file(s)
x, y, time, case_id = Di(fn_data).get_data(return_identifiers=True,
return_split_values=True)
print('\tM^{}_j ... '.format(s), end='', flush=True)
# S predictions
for i in sorted([int(i) for i in Name_functions.S_J_values(s)],
reverse=True):
if Filefunctions.exists(Name_functions.DSJ_probabilities(s, i)):
continue
model_i = Model_Functions.loadModel(Name_functions.model_SJ(s, i))
model_labels = model_i.classes_.tolist()
model_end_time = Name_functions.SJ_period_end_time(s, i)
with open(Name_functions.DSJ_probabilities(s, i), 'w+') as wf:
for dx, t, idn in zip(x, time, case_id):
if t < model_end_time:
# Only test if the model existed before the data point
continue
model_predictions = model_i.predict_proba(dx.reshape(1, -1))[0]
actual_predictions = [
(0 if (i not in model_labels) else
model_predictions[model_labels.index(i)])
for i in all_labels
]
wf.write('{};{};{}\n'.format(
idn, t,
';'.join(['{:4f}'.format(x) for x in actual_predictions])))
print('Done')
# Naive predictions
print('\tM^{}_naive ... '.format(s), end='', flush=True)
if Filefunctions.exists(Name_functions.DS_probabilities_naive(s)):
print('Already done')
return
model_naive = Model_Functions.loadModel(Name_functions.model_S_naive(s))
model_naive_labels = model_naive.classes_.tolist()
model_naive_end_time = Parameters.train_time_naive_stop
with open(Name_functions.DS_probabilities_naive(s), 'w+') as wf:
for dx, t, idn in zip(x, time, case_id):
if t < model_naive_end_time:
# Only test if the model existed before the data point
continue
model_predictions = model_naive.predict_proba(dx.reshape(1, -1))[0]
actual_predictions = [
(0 if (i not in model_naive_labels) else
model_predictions[model_naive_labels.index(i)])
for i in all_labels
]
wf.write('{};{};{}\n'.format(
idn, t,
';'.join(['{:4f}'.format(x) for x in actual_predictions])))
print('Done')
def create_labelled_dataset(event_log, s, feature_filename, output_file):
# Check or create Event Log
assert (isinstance(s, int))
if isinstance(event_log, EventLog):
pass
elif isinstance(event_log, str):
event_log = EventLog(filename=event_log)
else:
raise Exception('Given argument should be filename or EventLog')
# Dictionary with all labels
labels = dict()
# Each split of the full dataset gets parsed separately
splits = event_log.get_splits(s)
for split in splits.values():
labels.update(algorithm(split))
# Create labelled data
with open(feature_filename, 'r') as rf:
Filefunctions.makeParentDir(output_file)
with open(output_file, 'w+') as wf:
wf.write(rf.readline()[:-1] + ';Case_start;Class\n')
wf.write(rf.readline()[:-1] + ';-;-\n')
wf.write(rf.readline()[:-1] + ';-;-\n')
wf.write(rf.readline()[:-1] + ';SPLIT;Y\n')
for line in rf.readlines():
values = line[:-1].split(';')
case = event_log.get_case(values[0])
values.append(str(case.get_start()))
values.append(labels[values[0]])
wf.write(';'.join(values) + '\n')
def parse_ms(s):
print('D^{} ... '.format(s), end='', flush=True)
if Filefunctions.exists(Name_functions.DS_train_ids(s)):
if Filefunctions.exists(Name_functions.DS_test_ids(s)):
print('Already done')
return
np.random.seed(0)
X, y, times, ids = DI(Name_functions.DS_file(s)).get_data(
Name_functions.DS_reduced_ids_DSJ(s), True, True)
if Parameters.take_test_split_chronological:
test_case_ids = []
train_case_ids = []
times_post_warm_up = [
t for t in times if t > Parameters.test_time_start
]
times_post_warm_up.sort()
train_start_index = int(
(1 - Parameters.assessment_test_split) * len(times_post_warm_up))
train_time_end = times_post_warm_up[train_start_index]
for case_start_time, case_id in zip(times, ids):
if case_start_time <= Parameters.test_time_start:
continue
if case_start_time < train_time_end:
train_case_ids.append(case_id)
else:
test_case_ids.append(case_id)
else:
indices = [
i for i in range(len(ids)) if times[i] > Parameters.test_time_start
]
test_indices = []
train_indices = []
c, cc = np.unique(y[indices], return_counts=True)
for label, label_count in zip(c, cc):
num_test = int(label_count * Parameters.assessment_test_split)
indices_c = [i for i in indices if y[i] == label]
indices_c_test = np.random.choice(indices_c,
num_test,
replace=False)
test_indices.extend(indices_c_test.tolist())
train_indices.extend(
[i for i in indices_c if i not in indices_c_test])
test_case_ids = ids[test_indices]
train_case_ids = ids[train_indices]
with open(Name_functions.DS_train_ids(s), 'w+') as wf:
for case_id in train_case_ids:
wf.write('{}\n'.format(case_id))
with open(Name_functions.DS_test_ids(s), 'w+') as wf:
for case_id in test_case_ids:
wf.write('{}\n'.format(case_id))
print('Done')
def run():
fn = Name_functions.full_GRAEC_table()
# Clear Test Scores file
with open(fn, 'w+') as wf:
wf.write('Beta;Tau;P;S;Score_type;Score_Value\n')
for S in Parameters.S_values:
Filefunctions.make_directory(Name_functions.S_score_folder(S))
print('S = {}'.format(S))
parse_graec(S)
parse_previous(S)
parse_naive(S)
def drop_feature(fn, feature, fn_out=None):
Filefunctions.exists_assert(fn)
assert (isinstance(feature, str) or isinstance(feature, int))
df = import_df(fn)
if isinstance(feature, int):
feature = df.columns[feature]
if feature not in df.columns:
warn('Feature is not a column')
return
df.drop(labels=[feature], axis=1, inplace=True)
if fn_out is None:
fn_out = fn
export_df(df, fn_out)
def _eval_previous(self):
print('Parsing Previous ... ', end='', flush=True)
fn_recent = Name_functions.parameter_evaluation_evaluation_metric_file('Previous')
if Filefunctions.exists(fn_recent):
print('Already done')
return
with open(fn_recent, 'w+') as wf:
wf.write('S;Day;NumEntries;accuracy;f1\n')
for S in self.Multi['S']:
predictor = Classifiers.PreviousClassifier(S)
fn = Name_functions.DS_file(S)
_, labels, times, ids = Di(fn).get_data(fn_subset_ids=self.test_ids_fn,
return_split_values=True,
return_identifiers=True)
data = pd.DataFrame(index=ids)
data['time'] = times
data['y_true'] = [l[0] for l in labels]
data['Day'] = np.floor(data['time'])
# Calculate the accuracy score for each day
for day in data['Day'].unique():
subset = data[data['Day'] == day]
acc_score, f1_score = self.get_scores(predictor=predictor,
true_labels=subset['y_true'],
times=subset['time'],
ids=subset.index
)
if not (acc_score is None or f1_score is None):
wf.write('{};{};{};{};{}\n'.format(S,
day,
len(subset),
acc_score,
f1_score))
print('Done')
def parse_naive(s):
fn_target = Name_functions.DS_reduced_ids_naive(s)
print('\tD^S_naive ... ', end='', flush=True)
# Check existence
if Filefunctions.exists(fn_target):
print('Already done')
return
fn_input = Name_functions.DS_file(s)
x, y, timestamps, ids = DataImporter(fn_input).get_data(
return_identifiers=True, return_split_values=True)
first_year_indices = [
i for i in range(len(timestamps))
if timestamps[i] < Parameters.train_time_naive_stop
]
x = x[first_year_indices]
y = y[first_year_indices]
ids = ids[first_year_indices]
x, y, medoid_indices = KMedoids.reduce_to_medoids(
x, y, return_indices=True, factor=Parameters.LargeSmallFactor)
ids_keep = [ids[i] for i in medoid_indices]
with open(fn_target, 'w+') as wf:
for CaseID in ids_keep:
wf.write('{}\n'.format(CaseID))
print('Done')
def load_dict_from_csv(filename):
assert (Filefunctions.exists(filename))
ret = dict()
with open(filename, 'r') as rf:
for line in rf.readlines():
k, v = line[:-1].split(';', 1)
ret[k] = v
return ret
def sort_file_on_date(fn_in, feature, fn_out=None):
assert (isinstance(feature, str) or isinstance(
feature, int)), 'Feature must be str or int: {}'.format(feature)
Filefunctions.exists_assert(fn_in)
df = import_df(fn_in)
if isinstance(feature, int):
feature = df.columns[feature]
fixDates(frame=df, date_column_string=feature)
df.sort_values(by=[feature], ascending=True, inplace=True)
if fn_out is None:
fn_out = fn_in
export_df(df, fn_out)
def __init__(self, filename):
assert Filefunctions.exists(filename)
# We implement the event log as a dict, this allows easier reference when adding an event to the case
# Since each event has a reference to a case_id, not to a case itself
self.cases = dict()
with open(filename, 'r') as rf:
for line in rf:
# for each line
case_id, timestamp, act = line[:-1].split(';')
# create an event
e = Event(case_id=case_id, time=float(timestamp), act=act)
# add it to the corresponding case (or create the new case if necessary)
self.cases.setdefault(case_id,
Case(case_id=case_id)).add_event(e)
def split_data(self,
interval,
fn_subset_ids=None,
return_split_values=False,
return_identifiers=False):
if fn_subset_ids is not None:
if not Filefunctions.exists(fn_subset_ids):
raise Exception(
'File does not exist:\n{}'.format(fn_subset_ids))
keep_idx = None
if fn_subset_ids is not None:
with open(fn_subset_ids, 'r') as rf:
keep_ids = [line[:-1] for line in rf.readlines()]
keep_idx = [
i for i in range(len(self.IDS)) if self.IDS[i] in keep_ids
]
index_list = dict()
for (i, s) in enumerate(self.split_values):
if keep_idx is not None and i not in keep_idx:
continue
group = math.floor(float(s) / interval)
index_list.setdefault(group, []).append(i)
return_x = dict()
return_y = dict()
return_split = dict()
return_ids = dict()
for (group, indices) in index_list.items():
return_x[group] = self.X.toarray()[indices]
return_y[group] = self.y[indices]
if return_split_values:
return_split[group] = self.split_values[indices]
if return_identifiers:
return_ids[group] = self.IDS[indices]
ret = (
return_x,
return_y,
)
ret += ((return_split, ) if return_split_values else ())
ret += ((return_ids, ) if return_identifiers else ())
return ret
def _eval_param(self, evaluated_parameter):
print('Parsing parameter {} ... '.format(evaluated_parameter), end='', flush=True)
fn = Name_functions.parameter_evaluation_evaluation_metric_file(evaluated_parameter)
if Filefunctions.exists(fn):
print('Already done')
return
with open(fn, 'w+') as wf:
wf.write('S;Beta;Tau;P;Day;NumEntries;accuracy;f1\n')
for S in self.values(evaluated_parameter, 'S'):
predictor = Classifiers.BPTSClassifier(s=S, score_function=None)
fn = Name_functions.DS_file(S)
_, labels, times, ids = Di(fn).get_data(fn_subset_ids=self.test_ids_fn, return_split_values=True,
return_identifiers=True)
data = pd.DataFrame(index=ids)
data['time'] = times
data['y_true'] = [l[0] for l in labels]
data['Day'] = np.floor(data['time'])
for beta in self.values(evaluated_parameter, 'Beta'):
for p in self.values(evaluated_parameter, 'P'):
for tau in self.values(evaluated_parameter, 'Tau'):
scoring_function = PeriodScoring(s=S, beta=beta, tau=tau, p=p)
predictor.set_scoring_function(scoring_function)
for day in data['Day'].unique():
subset = data[data['Day'] == day]
acc_score, f1_score = self.get_scores(predictor=predictor,
ids=subset.index,
times=subset['time'],
true_labels=subset['y_true'],
)
if not (acc_score is None or f1_score is None):
wf.write('{};{};{};{};{};{};{};{}\n'.format(S,
beta,
tau,
p,
day,
len(subset),
acc_score,
f1_score))
print('Done')
def parse_ms(s):
fn_target = Name_functions.DS_reduced_ids_DSJ(s)
# Check existence
print('\tD^S_j ... ', end='', flush=True)
if Filefunctions.exists(fn_target):
print('Already done')
return
fn_input = Name_functions.DS_file(s)
x, y, ids = DataImporter(fn_input).split_data(int(s),
return_identifiers=True)
ids_keep = []
for i in sorted(x):
xi, yi, indices = KMedoids.reduce_to_medoids(x[i],
y[i],
return_indices=True)
ids_keep.extend([ids[i][j] for j in indices])
with open(fn_target, 'w+') as wf:
for caseID in ids_keep:
wf.write('{}\n'.format(caseID))
print('Done')
def get_data(self,
fn_subset_ids=None,
return_split_values=False,
return_identifiers=False):
if fn_subset_ids is None:
ret = (self.X.toarray(), self.y)
ret += (self.split_values, ) if return_split_values else ()
ret += (self.IDS, ) if return_identifiers else ()
else:
if not Filefunctions.exists(fn_subset_ids):
raise Exception(
'File does not exist:\n{}'.format(fn_subset_ids))
with open(fn_subset_ids, 'r') as rf:
keep_ids = [line[:-1] for line in rf.readlines()]
keep_idx = [
i for i in range(len(self.IDS)) if self.IDS[i] in keep_ids
]
ret = (self.X.toarray()[keep_idx], self.y[keep_idx])
ret += (
self.split_values[keep_idx], ) if return_split_values else ()
ret += (self.IDS[keep_idx], ) if return_identifiers else ()
return ret
# This script runs all other scripts in the project. MS stands for Moment / S, since experiments differ in the moment in
# in the process when the prediction takes place, and the subset period length S
def print_state(s):
print('-' * 25)
print(s)
print('-' * 25)
if Parameters.Demo:
print('Running Demo')
print_state('Building Dataset')
# Remove old folder
Filefunctions.delete(Parameters.root_location)
# Create new folder
Filefunctions.make_directory(Parameters.root_location)
DEMO_CREATE_EVENTLOG.run()
else:
print('Running Real Dataset')
print_state('Splitting and labeling event log')
STEP0_Split_And_Label.run()
# Reduce the size of the data set if needed
print_state('Reducing Data Size')
STEP1_Reducing_Class_Sizes.run()
# Train all models
print_state('Training')
def save_dict_to_csv(dictionary, filename):
assert (isinstance(dictionary, dict))
Filefunctions.makeParentDir(filename)
with open(filename, 'w+') as wf:
for k, v in dictionary.items():
wf.write('{};{}\n'.format(k, v))
def undo():
for S in Parameters.S_values:
Filefunctions.delete(Name_functions.S_score_folder(S))
STEP7_Global_Results.undo()
def undo():
Filefunctions.delete(Name_functions.results_folder())
STEP8_Over_Time_Scoring.undo()
def metric_figure(metric):
fn = results_folder() + '/Result_{}.png'.format(metric)
Filefunctions.makeParentDir(fn)
return fn
def parameter_evaluation_figure(parameter, metric):
fn = parameter_metric_evaluation_folder(metric) + '/{}.png'.format(
parameter)
Filefunctions.makeParentDir(fn)
return fn
def best_graec():
fn = parameter_evaluation_folder() + '/best_graec_parameters.csv'
Filefunctions.makeParentDir(fn)
return fn
def parse_ms(s):
print('\tGRAEC ... ', end='', flush=True)
if Filefunctions.exists(Name_functions.S_GRAEC_enumeration_dictionary(s)):
print('Already done')
return
enumeration_encoder = dict()
fn_data = Name_functions.DS_file(s)
fn_train_ids = Name_functions.DS_train_ids(s)
fn_test_ids = Name_functions.DS_test_ids(s)
x_train, labels_train, times_train, ids_train = DI(fn_data).get_data(
fn_subset_ids=fn_train_ids,
return_split_values=True,
return_identifiers=True)
x_test, labels_test, times_test, ids_test = DI(fn_data).get_data(
fn_subset_ids=fn_test_ids,
return_split_values=True,
return_identifiers=True)
enumeration = 0
predictor = Classifiers.BPTSClassifier(s=s, score_function=None)
for B in Parameters.GRAEC_beta:
for T in Parameters.GRAEC_tau:
for P in Parameters.GRAEC_p if not T == 0 else [
0
]: # P has no use for T == 0
enumeration_encoder[enumeration] = '{};{};{}'.format(B, T, P)
predictor.set_scoring_function(
score_function=PeriodScoring(beta=B, p=P, tau=T, s=s))
with open(
Name_functions.S_GRAEC_train_predictions(
s, enumeration), 'w+') as wf:
wf.write('SOID;time;True_label;Predicted_label\n')
for case_id, t, true_label in zip(ids_train, times_train,
labels_train):
predicted_label = predictor.predict(case_id=case_id,
time=t)
wf.write('{};{};{};{}\n'.format(
case_id, t, true_label[0], predicted_label))
with open(
Name_functions.S_GRAEC_test_predictions(
s, enumeration), 'w+') as wf:
wf.write('Case_id;time;True_label;Predicted_label\n')
for case_id, t, true_label in zip(ids_test, times_test,
labels_test):
predicted_label = predictor.predict(case_id=case_id,
time=t)
wf.write('{};{};{};{}\n'.format(
case_id, t, true_label[0], predicted_label))
enumeration += 1
Human_Functions.save_dict_to_csv(
enumeration_encoder, Name_functions.S_GRAEC_enumeration_dictionary(s))
fn_data = Name_functions.DS_file(s)
fn_ids = Name_functions.DS_test_ids(s)
x, labels, times, ids = DI(fn_data).get_data(fn_subset_ids=fn_ids,
return_split_values=True,
return_identifiers=True)
print('Done')
print('\tNaive and Previous ... ', end='', flush=True)
naive_predictor = Classifiers.NaiveClassifier(s)
previous_predictor = Classifiers.PreviousClassifier(s)
with open(Name_functions.S_naive_test_predictions(s), 'w+') as wf_naive:
with open(Name_functions.S_recent_test_predictions(s),
'w+') as wf_previous:
wf_naive.write('{};{};{};{}\n'.format('Case_id', 'time',
'True_label',
'Predicted_label'))
wf_previous.write('{};{};{};{}\n'.format('Case_id', 'time',
'True_label',
'Predicted_label'))
for case_id, t, true_label in zip(ids, times, labels):
predicted_label_naive = naive_predictor.predict(
case_id=case_id, time=t)
if predicted_label_naive is not None:
wf_naive.write('{};{};{};{}\n'.format(
case_id, t, true_label[0], predicted_label_naive))
predicted_label_previous = previous_predictor.predict(
case_id=case_id, time=t)
if predicted_label_previous is not None:
wf_previous.write('{};{};{};{}\n'.format(
case_id, t, true_label[0], predicted_label_previous))
print('Done')
def full_GRAEC_table():
fn = results_folder() + '/GRAEC Full Scores.csv'
Filefunctions.makeParentDir(fn)
return fn
def metric_table(metric):
fn = results_folder() + '/GRAEC TexTable {}.txt'.format(metric)
Filefunctions.makeParentDir(fn)
return fn
def S_J_values(S):
folder = S_folder(S) + '/Splits'
if Filefunctions.exists(folder):
return os.listdir(folder)
else:
return []
def parameter_evaluation_evaluation_metric_file(parameter):
fn = parameter_evaluation_data_folder() + '/{}.csv'.format(parameter)
Filefunctions.makeParentDir(fn)
return fn
def S_GRAEC_enumeration_folder(S):
fd = S_folder(S) + '/BEPT_Enumeration'
Filefunctions.make_directory(fd)
return fd
def undo():
for S in Parameters.S_values:
Filefunctions.delete(Name_functions.S_GRAEC_enumeration_folder(S))
Filefunctions.delete(Name_functions.S_naive_test_predictions(S))
Filefunctions.delete(Name_functions.S_recent_test_predictions(S))
STEP6_Global_Scoring.undo()
def undo():
Filefunctions.delete(Name_functions.parameter_evaluation_data_folder())
STEP9_Over_Time_Results.undo()
