def test_backward_subset_feature_selection(train_df, train_fn, eval_fn,
split_fn, base_extractor,
metric_name):
features_sets = {
"first": ["x1", "x2"],
"second": ["x4", "x5"],
"third": ["x3", "x6"]
}
logs = backward_subset_feature_selection(train_df,
train_fn,
features_sets,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=-1,
early_stop=10,
iter_limit=50,
min_remaining_features=5)
assert len(get_used_features(
first(logs)[0])) <= 5 # Assert stop by remaining features
logs = backward_subset_feature_selection(train_df,
train_fn,
features_sets,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=0,
early_stop=10,
iter_limit=1,
min_remaining_features=3)
assert len(logs) == 1 # Assert stop by iter limit
logs = backward_subset_feature_selection(train_df,
train_fn,
features_sets,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=1,
early_stop=2,
iter_limit=50,
min_remaining_features=1)
assert len(logs) == 2 # Assert stop by early_stop
def test_poor_man_boruta_selection(train_df, holdout_df, train_fn, eval_fn,
base_extractor, metric_name):
features = ["x1", "x2", "x3", "x4", "x5", "x6"]
logs = poor_man_boruta_selection(train_df,
holdout_df,
train_fn,
features,
eval_fn,
base_extractor,
metric_name,
max_removed_by_step=1,
threshold=0,
early_stop=10,
iter_limit=50,
min_remaining_features=5)
assert len(get_used_features(
first(logs))) <= 6 # Assert stop by remaining features
logs = poor_man_boruta_selection(train_df,
holdout_df,
train_fn,
features,
eval_fn,
base_extractor,
metric_name,
max_removed_by_step=1,
threshold=0,
early_stop=10,
iter_limit=1,
min_remaining_features=3)
assert len(logs) == 1 # Assert stop by iter limit
logs = poor_man_boruta_selection(train_df,
holdout_df,
train_fn,
features,
eval_fn,
base_extractor,
metric_name,
max_removed_by_step=1,
threshold=1,
early_stop=2,
iter_limit=50,
min_remaining_features=1)
assert len(logs) == 2 # Assert stop by early_stop
def test_feature_importance_backward_selection(train_df, train_fn, eval_fn,
split_fn, base_extractor,
metric_name):
features = ["x1", "x2", "x3", "x4", "x5", "x6"]
logs = feature_importance_backward_selection(train_df,
train_fn,
features,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=0,
early_stop=10,
iter_limit=50,
min_remaining_features=5)
assert len(get_used_features(
first(logs))) <= 5 # Assert stop by remaining features
logs = feature_importance_backward_selection(train_df,
train_fn,
features,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=0,
early_stop=10,
iter_limit=1,
min_remaining_features=3)
assert len(logs) == 1 # Assert stop by iter limit
logs = feature_importance_backward_selection(train_df,
train_fn,
features,
split_fn,
eval_fn,
base_extractor,
metric_name,
num_removed_by_step=1,
threshold=1,
early_stop=2,
iter_limit=50,
min_remaining_features=1)
assert len(logs) == 2 # Assert stop by early_stop
def stop_by_num_features(logs: ListLogListType,
min_num_features: int = 50) -> bool:
"""
Checks for logs to see if feature selection should stop
Parameters
----------
logs : list of list of dict
A list of log-like lists of dictionaries evaluations.
min_num_features: int (default 50)
The minimun number of features the model can have before stopping
Returns
-------
stop: bool
A boolean whether to stop recursion or not
"""
return len(get_used_features(first(logs))) <= min_num_features
def test_get_used_features(logs):
result = get_used_features(logs[0])
assert result == ['x1', 'x2', 'x4', 'x5', 'x3', 'x6']
def remove_by_feature_shuffling(log: LogType,
predict_fn: PredictFnType,
eval_fn: EvalFnType,
eval_data: pd.DataFrame,
extractor: ExtractorFnType,
metric_name: str,
max_removed_by_step: int = 50,
threshold: float = 0.005,
speed_up_by_importance: bool = False,
parallel: bool = False,
nthread: int = 1,
seed: int = 7) -> List[str]:
"""
Performs feature selection based on the evaluation of the test vs the
evaluation of the test with randomly shuffled features
Parameters
----------
log : LogType
Dictionaries evaluations.
predict_fn: function pandas.DataFrame -> pandas.DataFrame
A partially defined predictor that takes a DataFrame and returns the
predicted score for this dataframe
eval_fn : function DataFrame -> log dict
A partially defined evaluation function that takes a dataset with prediction and
returns the evaluation logs.
eval_data: pandas.DataFrame
Data used to evaluate the model after shuffling
extractor: function str -> float
A extractor that take a string and returns the value of that string on a dict
metric_name: str
String with the name of the column that refers to the metric column to be extracted
max_removed_by_step: int (default 5)
The maximum number of features to remove. It will only consider the least max_removed_by_step in terms of
feature importance. If speed_up_by_importance=True it will first filter the least relevant feature an
shuffle only those. If speed_up_by_importance=False it will shuffle all features and drop the last
max_removed_by_step in terms of PIMP. In both cases, the features will only be removed if drop in
performance is up to the defined threshold.
threshold: float (default 0.005)
Threshold for model performance comparison
speed_up_by_importance: bool (default True)
If it should narrow search looking at feature importance first before getting PIMP importance. If True,
will only shuffle the top num_removed_by_step in terms of feature importance.
parallel: bool (default False)
nthread: int (default 1)
seed: int (default 7)
Random seed
Returns
----------
features: list of str
The remaining features after removing based on feature importance
"""
random.seed(seed)
curr_metric = get_avg_metric_from_extractor(log, extractor, metric_name)
eval_size = eval_data.shape[0]
features_to_shuffle = order_feature_importance_avg_from_logs(log)[-max_removed_by_step:] \
if speed_up_by_importance else get_used_features(log)
def shuffle(feature: str) -> pd.DataFrame:
return eval_data.assign(
**{feature: eval_data[feature].sample(frac=1.0)})
feature_to_delta_metric = compose(
lambda m: curr_metric - m,
get_avg_metric_from_extractor(extractor=extractor,
metric_name=metric_name),
gen_validator_log(fold_num=0, test_size=eval_size), eval_fn,
predict_fn, shuffle)
if parallel:
metrics = Parallel(n_jobs=nthread, backend="threading")(
delayed(feature_to_delta_metric)(feature)
for feature in features_to_shuffle)
feature_to_delta_metric = dict(zip(features_to_shuffle, metrics))
gc.collect()
else:
feature_to_delta_metric = {
feature: feature_to_delta_metric(feature)
for feature in features_to_shuffle
}
return pipe(feature_to_delta_metric,
valfilter(lambda delta_metric: delta_metric < threshold),
sorted(key=lambda f: feature_to_delta_metric.get(f)),
take(max_removed_by_step), list)
