Xgbfi

XGBoost Feature Interactions & Importance

What is Xgbfi?

Xgbfi is a XGBoost model dump parser, which ranks features as well as feature interactions by different metrics.

Siblings

Xgbfir - Python porting

The Metrics

• Gain: Total gain of each feature or feature interaction
• FScore: Amount of possible splits taken on a feature or feature interaction
• wFScore: Amount of possible splits taken on a feature or feature interaction weighted by the probability of the splits to take place
• Average wFScore: wFScore divided by FScore
• Average Gain: Gain divided by FScore
• Expected Gain: Total gain of each feature or feature interaction weighted by the probability to gather the gain
• Average Tree Index
• Average Tree Depth

Additional Features

• Leaf Statistics
• Split Value Histograms

Example:

Usage

[mono] XgbFeatureInteractions.exe [-help|options]

Quick Guide

a) Creating a feature map (fmap)

def create_feature_map(fmap_filename, features):
"""
features: enumerable of feature names
"""
    outfile = open(fmap_filename, 'w')
    for i, feat in enumerate(features):
        outfile.write('{0}\t{1}\tq\n'.format(i, feat))
    outfile.close()

create_feature_map('xgb.fmap', features) 

b) Dumping a XGBoost model

gbdt.dump_model('xgb.dump',fmap='xgb.fmap', with_stats=True)

c) Editing Parameters in XgbFeatureInteractions.exe.config

<setting name="XgbModelFile" serializeAs="String">
    <value>xgb.dump</value>
</setting>

d) Running [mono] XgbFeatureInteractions.exe without cmd line parameters