how can make a process after dimension reduction like ICA ?

norah

hello,
I want to ask about dimension reduction I used ICA on dataset about climate it has (temperature degree, pressure and so on) so after reducing dimension I get attribute with name IC1, IC2 and so on.
now I want to make some operations on it but I confused how can I make a process like summation if I want on temperature even I don't temperature attribute name now.
depend on what can I make the process after dimension reduction?

varunm1

Hello @norah

When you apply dimension reduction techniques like ICA each independent component formed is based on the weights from all features in our dataset. So there is no direct relation between Independent component and your original feature as the new IC's are based on all and represent a different thing compared to features. Once you apply ICA on a dataset it tries to tell the source of your features rather than the features itself which is the reason this is called blind source separation.

varunm1

I will explain to you a famous use case. Blind source separation techniques like ICA are used for finding hidden variables.

Example: Imagine that there are four people in a room talking at a time with each other and you have two microphones that record their audio data. Now you have two audio signals that consist of all 4 people voices in each one of them but it is difficult for you to identify people based on audio data (mixed). Now what ICA does is it takes different features from audio signals and create four independent components so that each belongs to a particular person. The main use of ICA technique is to find independent hidden variables. 

Based on the math behind this algorithm we can use it in dimensionality reduction.

Thanks 

varunm1

Hello @norah

When you apply dimension reduction techniques like ICA each independent component formed is based on the weights from all features in our dataset. So there is no direct relation between Independent component and your original feature as the new IC's are based on all and represent a different thing compared to features. Once you apply ICA on a dataset it tries to tell the source of your features rather than the features itself which is the reason this is called blind source separation.

norah

ok, what's the benefit from this feature I mean dimension reduction except for low space and time consuming if I can not know the original attributes? can you give me an example of using this 

varunm1

I will explain to you a famous use case. Blind source separation techniques like ICA are used for finding hidden variables.

Example: Imagine that there are four people in a room talking at a time with each other and you have two microphones that record their audio data. Now you have two audio signals that consist of all 4 people voices in each one of them but it is difficult for you to identify people based on audio data (mixed). Now what ICA does is it takes different features from audio signals and create four independent components so that each belongs to a particular person. The main use of ICA technique is to find independent hidden variables. 

Based on the math behind this algorithm we can use it in dimensionality reduction.

Thanks 

norah

I got it, thank you so much ..