Python implementation of TBFS feature ranking algorithm.
pip install git+https://github.com/johnsonj561/Threshold-Based-Feature-Selection
See notebook for example usage.
from tbfs.ranker import TBFSRanker, metrics
import pandas as pd
# prepare data features and labels
df = pd.read_csv('sample-data.csv')
y, x = df['class'], df.drop(columns=['class'])
y = np.where(y == 'ACL', 1, 0)
# fit TBFS ranker
tbfs = TBFSRanker(t_delta=0.01)
tbfs.fit(x, y)
# take top K features for a metric
tbfs.top_k_features_by_metric(10, 'f-score')['GENE1609X',
'GENE1537X',
'GENE493X',
'GENE1616X',
'GENE3945X',
'GENE3258X',
'GENE3946X',
'GENE384X',
'GENE1296X',
'GENE1620X']
We can save feature rankings for future experiments and re-use them as needed.
# save results
tbfs.to_csv('tbfs-results.csv')
...
# load results and re-use
tbfs2 = TBFSRanker()
tbfs2.from_csv('tbfs-results.csv')
tbfs2.top_k_features_by_metric(10, 'f-score')['GENE1609X',
'GENE1537X',
'GENE493X',
'GENE1616X',
'GENE3945X',
'GENE3258X',
'GENE3946X',
'GENE384X',
'GENE1296X',
'GENE1620X']
When working with highly imbalanced data, we often want to explore very small thresholds.
For example, a data set with minority class size of 0.01% generally requires a threshold of 0.0001.
Setting t_delta to 0.0001 would cause the TBFS ranker to enumerate 10,000 thresholds and will drastically increase run time.
We can address this by providing a custom threshold range that varies t_delta. The below example reduces 10,000 thresholds to < 500 thresholds.
thresholds = [
*np.arange(0, 0.01, 0.0001),
*np.arange(0.01, 0.1, 0.001),
*np.arange(0.1, 0.9, 0.01),
*np.arange(0.9, 0.99, 0.001),
*np.arange(0.99, 1.0001, 0.0001)
]
tbfs = TBFSRanker(thresholds=)
tbfs.fit(x,y)
...