Scikit-Learn Pipeline Integration

The feature generation process of this package follows a certain workflow; first, the connection to the knowledge graph entities is established (see Linking), then features are generated (see Generators). Optionally, more links can be obtained by using the Link Exploration. For the subsequent filtering of the additional features, the Hierarchical Feature Selection algorithms can be used. In case of URI features referring to the same entity, Schema Matching & Fusion can be applied.

For convenience, this workflow can be applied within a scikit-learn pipeline. The following example shows how to set up the pipeline in order to generate and filter features for this example dataframe:



Stephen King


Joanne K. Rowling


Dan Brown


import pandas as pd
from sklearn.pipeline import Pipeline

from kgextension.linking_sklearn import DbpediaLookupLinker
from kgextension.generator_sklearn import SpecificRelationGenerator, DirectTypeGenerator
from kgextension.feature_selection_sklearn import HierarchyBasedFilter

# input DataFrame
df = pd.DataFrame({
    'author': ['Stephen King', 'Joanne K. Rowling', 'Dan Brown'],
    'book_sales': [20, 25, 18]

# set up the parts of the pipeline
Linker = DbpediaLookupLinker(column='author')
Generator1 = SpecificRelationGenerator(columns=['new_link'])
Generator2 = DirectTypeGenerator(columns=['new_link'], hierarchy=True)
Filter = HierarchyBasedFilter(label_column='book_sales')

# combine the functions into the pipeline
pipeline = Pipeline(steps = [('lookup_linker', Linker),
                             ('sr_generator', Generator1),
                             ('dt_generator', Generator2),
                             ('hb_filter', Filter)])

# fit and transform the input data