def get_xgboost_model(name="xgboost_common"):
return scope.get_xgb_model(
n_estimators=scope.int(
hp.quniform(
get_full_name(name, "n_estimators"),
1, 200, 1,
),
),
max_depth=scope.int(
hp.quniform(
get_full_name(name, 'max_depth'),
1, 13, 1,
),
),
min_child_weight=scope.int(
hp.quniform(
get_full_name(name, 'min_child_weight'),
1, 6, 1,
),
),
subsample=scope.int(
hp.uniform(
get_full_name(name, 'subsample'),
0.5, 1,
),
),
gamma=hp.uniform(
get_full_name(name, 'gamma'),
0.5, 1,
),
nthread=1,
seed=RANDOM_STATE,
)
def get_model_params(name="model_common", xgb_name=None, lr_name=None):
xgb_result_name = xgb_name if xgb_name is not None else get_full_name(
name, 'xgb')
lr_result_name = lr_name if lr_name is not None else get_full_name(
name, 'lr')
return hp.choice(name, (
get_xgboost_params(xgb_result_name),
get_linear_model_params(lr_result_name),
get_rf_model_params('rf'),
))
def get_linear_model(name="logistic_regression"):
"""Set up hyperparameters and ranges for logistic regression model."""
return scope.get_lr_model_wrapper(
C=hp.uniform(get_full_name(name, 'C'), 0, 1),
penalty=hp.choice(get_full_name(name, 'penalty'), ('l1', 'l2')),
class_weight=hp.choice(get_full_name(name, 'class_weight'),
(defaultdict(lambda: 1.0), 'balanced')),
fit_intercept=hp.choice(get_full_name(name, 'fit_intercept'),
(False, True)),
random_state=RANDOM_STATE)
def get_min_size_prefilter_params(name="min_size_prefilter_common"):
return scope.get_min_size_prefilter(min_size=hp.quniform(
get_full_name(name, 'min_size'),
0,
10,
1,
), )
def get_bayes_based_priority_getter_params(
name='bayes_based_priority_getter_common'):
return scope.get_bayes_based_priority_getter(max_features=hp.quniform(
get_full_name(name, 'max_features'),
0,
3000,
1,
), )
def get_min_simple_features_index_getter_params(
name="min_simple_features_index_getter_common"):
return scope.get_min_simple_features_index_getter(
use_raw_candidate=hp.choice(
get_full_name(name, 'use_raw_candidate'),
(False, True),
),
max_check=1000,
)
def get_priority_getter_params(name='priority_getter_common'):
return hp.choice(
name,
(
scope.get_simple_priority_getter(),
get_bayes_based_priority_getter_params(
get_full_name(name, 'bayes_priority_getter'), ),
),
)
def __init__(self, lower_level_model=None):
super(BasicModelSelectorPrimitive, self).__init__()
# set lower-level model
self._lower_level_model = lower_level_model
# set model name for logging
name = "model_common"
xgb_name = 'xgb'
lr_name = 'lr'
rf_name = 'rf'
# set model structure
self._model_structure = hp.choice(name, (
get_xgboost_params(name=get_full_name(name, xgb_name)),
get_linear_model(name=get_full_name(name, lr_name)),
get_rf_model_params(name=get_full_name(name, rf_name)),
))
def get_feature_selector_params(inner_model_params,
name='feature_selector_params'):
model_based_estimator = get_feature_selector_estimator_params()
return hp.choice(
name,
(
get_k_best_params(
inner_model_params=inner_model_params,
name=get_full_name(name, 'k_best'),
),
#get_boruta_feature_selector_params(
# name=get_full_name(name, 'boruta'),
# inner_model_params=inner_model_params,
#),
get_model_based_feature_selector_params(
name=get_full_name(name, 'model_based_selector'),
estimator=model_based_estimator,
inner_model_params=inner_model_params,
),
),
)
def __init__(self):
super(MultimarkerFeatureSelectorPrimitive, self).__init__()
# range for feature selectino threshold
l_threshold = -5
u_threshold = 5
name = 'multimarker_feature_selector'
xgb_name = 'xgb'
lr_name = 'lr'
rf_name = 'rf'
# structure of the current lever - we will use existing class from previous version from this
self._model_structure = scope.get_model_based_feature_selection_model(
# lower layer model for estimating features importance
estimator=hp.choice(name, (
get_rf_model_params(name=get_full_name(name, rf_name)),
get_linear_model(name=get_full_name(name, lr_name)),
)),
# current layer model that works on filtered features
inner_model=hp.choice(name, (
get_xgboost_params(name=get_full_name(name, xgb_name)),
get_linear_model(name=get_full_name(name, lr_name)),
get_rf_model_params(name=get_full_name(name, rf_name)),
)),
# threshold for filtering important features
feature_selection_threshold_coef=hp.loguniform(
get_full_name(name, "threshold"), l_threshold, u_threshold),
)
def __init__(self, lower_level_model):
super(SingelmarkerFeatureSelectorPrimitive, self).__init__()
name = 'single_marker_k_best_selector'
xgb_name = 'xgb'
lr_name = 'lr'
rf_name = 'rf'
self._lower_level_model = lower_level_model
self._model_structure = scope.get_k_best_wrapper(
# current layer model that works on filtered features
inner_model=hp.choice(name, (
get_xgboost_params(name=get_full_name(name, xgb_name)),
get_linear_model_params(name=get_full_name(name, lr_name)),
get_rf_model_params(name=get_full_name(name, rf_name)),
)),
# lower layer model for estimating features importance
k_best=scope.get_k_best(
k=hp.qloguniform(get_full_name(name, 'k'), 0, 5, 1),
score_func=hp.choice(get_full_name(name, 'score_func'),
(chi2, f_classif)),
),
)
def get_and_based_index_getter_params(name="and_based_index_getter_common"):
return scope.get_and_based_index_getter(use_raw_candidate=hp.choice(
get_full_name(name, 'use_raw_candidate'),
(False, True),
), )
def get_feature_selector_estimator_params(name='feature_selector_estimator'):
return hp.choice(name, (
get_rf_model_params(name=get_full_name(name, 'rf')),
get_linear_model_params(name=get_full_name(name, 'lr')),
))
