Altair RISE

A program to recognize and reward our most engaged community members

Nominate Yourself Now!

L2 Norm

Is there any way to perform an L2 Norm on feature vectors e.g. normalize the vectors to length 1.

The normalize operator seems to only scale within features not across features.

Thanks,
-Gagi

Find more posts tagged with

AI Studio

Accepted answers

All comments

land

Hi,
unfortunately this isn't currently possible with a built in operator. If you don't have too many attributes, you might make the calculations with the Generate Attributes operator.
Why are you going to normalize the complete example this way? The attributes will receive different weights per examples and learning will nearly be impossible.

Greetings,
Sebastian

dragoljub

Actually, depending on the learning task the L2 norm can be quite effective. For example say you have a long vector of features and want to check when only a few of many is abnormally different compared to the population trend. When using dot product kernels for example the L2-Norm can allow you to compute the angle between samples rather quickly (dot between normalized vector = cos(theta)). Also there is the added benefit of working with small feature magnitudes that speeds up computations in LibSVM for example. So thats why I'm using it. There are other examples, and I think it will be a great feature to add to the normalization operator.

-Gagi

Sebastian Land wrote:

Hi,
unfortunately this isn't currently possible with a built in operator. If you don't have too many attributes, you might make the calculations with the Generate Attributes operator.
Why are you going to normalize the complete example this way? The attributes will receive different weights per examples and learning will nearly be impossible.

Greetings,
Sebastian

land

Hi,
I thought about this yesterday evening when driving home and I came to exactly the same conclusion. I've added this on the growing stack of good ideas that are still to implement...

Greetings,
Sebastian

dragoljub

Great! :P