FRANe

Unsupervised Feature Ranking via Attribute Networks (FRANe) converts a dataset into a network (graph) with

• nodes that correspond to the features in the data,
• undirected edges whose weights are proportional to the similarity between the two corresponding features.

PageRank algorithm is than used to compute the centrality of the nodes (features) and the computed scores are interpreted as feature importance scores.

Examplary Code Snippet

The FRANe method is implemented in Python3. The implementation requires some standard scientific libraries (e.g., numpy and scipy) that make the implementation efficient.

The method is easy to use:

import frane
import numpy as np

x = np.random.random((100,1000))
r = frane.FRANe()
r.fit(x)
scores = r.feature_importances_
print(scores)

See examples for more examples. To run tests, please try pytest ./tests/*

Data

The data in the directory data was taken from sk-feature repository.