Skip to content

Latest commit

 

History

History
73 lines (55 loc) · 2.37 KB

Feature engineering.md

File metadata and controls

73 lines (55 loc) · 2.37 KB
Raw

Feature engineering

A critical part of machine learning is feature engineering. BlueCast's pipelines will automatically execute only necessary feature engineering and leaves this to the end user. However BlueCast offers some tools for feature engineering to make this part more approachable and faster.

First we import the required modules:

from bluecast.preprocessing.feature_types import FeatureTypeDetector
from bluecast.preprocessing.feature_creation import AddRowLevelAggFeatures, GroupLevelAggFeatures

Next we can make use of FeatureTypeDetector to identify numerical columns:

ignore_cols = [TARGET, "id", "CustomerId"]

feat_type_detector = FeatureTypeDetector()
train_data = feat_type_detector.fit_transform_feature_types(train.drop(ignore_cols, axis=1))

Next we use AddRowLevelAggFeatures to create features on row level. This usually adds a small degree of additional performance.

agg_feat_creator = AddRowLevelAggFeatures()

train_num = agg_feat_creator.add_row_level_agg_features(train.loc[:, feat_type_detector.num_columns])
test_num = agg_feat_creator.add_row_level_agg_features(test.loc[:, feat_type_detector.num_columns])

train_num = train_num.drop(agg_feat_creator.original_features, axis=1)
test_num = test_num.drop(agg_feat_creator.original_features, axis=1)


train = pd.concat([train, train_num], axis=1)
test = pd.concat([test, test_num], axis=1)

Additionally we can also provide information via group aggregations with GroupLevelAggFeatures:

group_agg_creator = GroupLevelAggFeatures()

train_num = group_agg_creator.create_groupby_agg_features(
    df = train,
    groupby_columns=["Geography", "Gender", "NumOfProducts"],
    columns_to_agg=feat_type_detector.num_columns, # None = take all
    target_col=None,
    aggregations = None # falls back to some aggs
)

test_num = group_agg_creator.create_groupby_agg_features(
    df = test,
    groupby_columns=["Geography", "Gender", "NumOfProducts"],
    columns_to_agg=feat_type_detector.num_columns, # None = take all
    target_col=TARGET,
    aggregations = None # falls back to some aggs
)

# joining the train information everywhere
train = train.merge(train_num, on=["Geography", "Gender", "NumOfProducts"], how="left")
test = test.merge(train_num, on=["Geography", "Gender", "NumOfProducts"], how="left")

Please note that this will increase the number of features significantly.