Do we need to normalize the Data in outlier detection?

[Anki]

Hello All,

For finding the outliers, Do we need to normalize the Data?
Can you please tell me impact without normalizing?
Which methods are best for Normalized data for Outlier detection?
And Which methods are better if I don't want to normalize?

Thank you in advance

Yours
Anki

[Ingo Mierswa]

Hi,

For finding the outliers, Do we need to normalize the Data?

this is certainly recommended, yes. All outlier detection methods depend on distances and / or densities and distance calculations can be really skewed on non-normalized data.

Can you please tell me impact without normalizing?

If you have a single dimension with a much larger scale than the other dimensions, this single dimension might overrule the others.

Which methods are best for Normalized data for Outlier detection?

Totally depends, there should probably also be a No-Free-Lunch-Theorem for outlier detection 
In general, I have good experiences with the local outlier factor method.

And Which methods are better if I don't want to normalize?

Don't do this (see above).

Cheers,
Ingo

[Anki]

Hi Ingo,

Thank you very much.

Yours
Anki

[wessel]

Agreed with all above.

I would like to add:
Visualization. Finding a way to nicely visualize your data can be time consuming but should be worth it.
You can often spot outliers yourself.

You can also look at misclassifications. The ada-boost algorithm adds weight to instances that are misclassified.
Outliers are often given a very high weight after only a few iterations.

Best regards,

Wessel