4
$\begingroup$

I have a dataset where I'd like to perform anomaly detection with an Isolation Forest. I don't have any way to validate the model (my data is not labeled - that's why I'm using unsupervised learning) - how can I tell if the model is working all right? I could do a train-test split, but again, how do I know if the predictions are correct if I'm using unlabelled data (plus I'd like to have as many words in my tf-idf vectoriser as possible, but that's another question)? There isn't any data about the amount of contamination either. How should I fine-tune the parameters/validate the results?

$\endgroup$
1

1 Answer 1

1
$\begingroup$

For isolation forest, here is a clue for validation reference.

From the paper and sklearn lib,we know there are two key parameters: n_estimators and max_samples.

  • when n_estimator = 100, the average path( score of outlier) is convergence.
  • when max_samples = 256(the default parameter), the different dataset will be convergence similar auc. that means no need to spend time for bigger samples.

According to the paper, they test difference n_estimator and max_sample to make sure the result is convergence.

So, for you case, you may need use grid search to check which combination of n_estimators and max_samples could be quick reach convergence.

In my case, I use titanic dataset. After scale raw data, the feature quantity is about 120. I used gridsearch and found n_estimator =200 and max_sample= 256, the outlier prediction begin to convergence.

This is my way to validate unsupervised outlier detection.

Hope it helpful.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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