IForest feature contribution for a given outlier

[lujiazho]

Hi, is it possible to print out which feature directly/more likely contributes to determining a data point as an outlier? So that we can check for each data point that which features may deviate the data point from the distribution the most.

[KulikDM]

Hi @lujiazho, you can actually use the shap library for this:

# Fit PyOD model
clf = KNN()
clf.fit(data)
    
# Shap is slow so perhaps only explain highest likelihood values, e.g. 100
scores = clf.decision_scores_
idx = np.argsort(scores)
    
# Fit shap explainer and get values for top 100
explainer = shap.Explainer(clf.decision_function, data)
shap_values = explainer(data[idx][-100:])

# Or for just one point of interest
idx = 55
shap_values = explainer(data[55].reshape(1,-1))
    
# Example of some plots
shap.plots.waterfall(shap_values)
shap.plots.beeswarm(shap_values)
shap.plots.bar(shap_values)

Hope this is what you were looking for and all the best!

[lujiazho]

Thanks a lot! That's exactly what I want!

[jesuinovieira]

You can also use clf.feature_importances_() for IForest