3
$\begingroup$

I am using Isolation forest for anomaly detection on multidimensional data. The algorithm is detecting anomalous records with good accuracy. Apart from detecting anomalous records I also need to find out which features are contributing the most for a data point to be anomalous. Is there any way we can get this?

$\endgroup$
5
$\begingroup$

SHAP values and the shap Python library can be used for this. Shap has built-in support for scikit-learn IsolationForest since October 2019.

import shap
from sklearn.ensemble import IsolationForest

# Load data and train Anomaly Detector as usual 
X_train, X_test, ...
est = IsolationForest()
est.fit(...)

# Create shap values and plot them
X_explain = X_test
shap_values = shap.TreeExplainer(est).shap_values(X_explain)
shap.summary_plot(shap_values, X_explain)

Here is an example of a plot I did for one IsolationForest model that I had, which was time-series. enter image description here

You can also get partial dependence plots for a particular feature, or a plot showing the feature contributions for a single X instance. Examples for this is given in the shap project README.

$\endgroup$
0
$\begingroup$

One possible describing feature importance in unsupervised outlier detecion is described in Contextual Outlier Interpretation. Similar as in the Lime approach, local linearity is assumed and by sampling a data points around the outlier of interest a classification problem is generated. The authors suggest to apply a SVM with linear kernel and use estimeited weights for feature importance.

$\endgroup$
1
  • $\begingroup$ I am wondering if you can use LIME to generate local explanations for each of the identified anomalies in IsolationForest? $\endgroup$ – FlyingPickle Oct 4 '20 at 14:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.