def _calc_metrics(idx_data, dataset, changepoints, tolerance_delay, threshold,
retrospective): # calculate the metrics
_retrospective = deepcopy(retrospective)
scores = _retrospective.calc_scores(dataset[idx_data])
F1_score, precision, recall = calc_F1_score(scores, changepoints,
tolerance_delay, threshold)
return F1_score, precision, recall
def _calc_AW2S_MDL_metrics(idx_data, dataset, changepoints, tolerance_delay,
threshold, retrospective):
_retrospective = deepcopy(retrospective)
alarms = _retrospective.make_alarms(dataset[idx_data])
scores = np.zeros(len(dataset[idx_data]))
scores[alarms] = 1
F1_score, precision, recall = calc_F1_score(scores, changepoints,
tolerance_delay, threshold)
return F1_score, precision, recall
def _objective_CF(trial, train, changepoints, tolerance_delay): # ChangeFinder
# hyperparameters
r = trial.suggest_uniform('r', 0.01, 0.99)
order = trial.suggest_int('order', 1, 20)
smooth = trial.suggest_int('smooth', 3, 20)
retrospective = changefinder.Retrospective(r=r, order=order, smooth=smooth)
scores = retrospective.calc_scores(train)
F1_score, _, _, _ = calc_F1_score(scores, changepoints, tolerance_delay)
return -F1_score
def _objective_BOCPD(trial, train, changepoints, tolerance_delay): # BOCPD
lam = trial.suggest_int('lam', 2, 1000)
alpha = trial.suggest_uniform('alpha', 0.01, 10)
beta = trial.suggest_uniform('beta', 0.01, 10)
kappa = trial.suggest_uniform('kappa', 0.01, 10)
mu = trial.suggest_uniform('mu', 0.01, 10)
h = partial(bocpd.constant_hazard, lam)
lik = bocpd.StudentT(alpha, beta, kappa, mu)
retrospective = bocpd.Retrospective(hazard_func=h, likelihood_func=lik)
scores = retrospective.calc_scores(train)
F1_score, _, _, _ = calc_F1_score(scores, changepoints, tolerance_delay)
return -F1_score
def _objective_AW2S_MDL(trial, train, changepoints, tolerance_delay):
window_size = trial.suggest_int('window_size', 10, 500)
sigma_given = trial.suggest_uniform('sigma_given', 0.1, 2)
nml_gaussian = partial(sdmdl_nml.nml_gaussian,
mu_max=1e8,
div_min=1e-8,
div_max=1e8)
complexity_gaussian = partial(sdmdl_nml.complexity_gaussian,
mu_max=1e8,
div_min=1e-8,
div_max=1e8)
retrospective_first = sdmdl.Retrospective(
h=window_size,
encoding_func=nml_gaussian,
complexity_func=complexity_gaussian,
order=0)
lnml_gaussian = partial(hsdmdl2_nml.lnml_gaussian, sigma_given=sigma_given)
retrospective_second = hsdmdl2.Retrospective(encoding_func=lnml_gaussian,
d=2,
M=5,
min_datapoints=5,
delta_0=0.05,
delta_1=0.05,
delta_2=0.05,
how_to_drop='all',
order=0,
reliability=True)
retrospective = aw2s_mdl.Retrospective(retrospective_first,
retrospective_second)
alarms = retrospective.make_alarms(train)
scores = np.zeros(len(train))
scores[alarms] = 1
F1_score, _, _ = calc_F1_score(scores,
changepoints,
tolerance_delay,
tuned_threshold=0.5)
return -F1_score
def _objective_Hierarchical(trial, train, changepoints, tolerance_delay,
order):
nml_gaussian = partial(hsdmdl2_nml.nml_gaussian)
min_datapoints = 5
# delta_0だけはいかなる場合でもチューニングしなければならない
delta_0 = trial.suggest_uniform('delta_0', 0.000001, 10.00)
if order == 1:
delta_1 = trial.suggest_uniform('delta_1', 0.000001, 1.00)
else:
delta_1 = 0.05
if order == 2:
delta_2 = trial.suggest_uniform('delta_2', 0.000001, 1.00)
else:
delta_2 = 0.05
retrospective = hsdmdl2.Retrospective(encoding_func=nml_gaussian,
d=2,
M=5,
min_datapoints=min_datapoints,
delta_0=delta_0,
delta_1=delta_1,
delta_2=delta_2,
how_to_drop='all',
order=order,
reliability=True)
alarms = retrospective.make_alarms(train)
scores = np.zeros(len(train))
scores[alarms] = 1
F1_score, _, _ = calc_F1_score(scores,
changepoints,
tolerance_delay,
tuned_threshold=0.5)
return -F1_score
def calc_opt_threshold(train, changepoints, tolerance_delay, retrospective):
_retrospective = deepcopy(retrospective)
scores = _retrospective.calc_scores(train)
_, _, _, opt_threshold = calc_F1_score(scores, changepoints,
tolerance_delay)
return opt_threshold