1
$\begingroup$

I'm trying to do Embedded clustering using kmeans.

This is customer data, so it involves a lot of sentences, so I'm using the universal sentence encoder before clustering.

But I should be doing a feature selection or dimensionality reduction before embedding the features.

I want to know if there is a method to do feature selection or dimensionality reduction in unsupervised learning. This could be very helpful as the clustering is giving a mixed result as of now and I have a strong feeling that this could be because of the unwanted attribute in the data.

I have read all the resources which only gave options to be done on the supervised learning.

Any help is appreciated!

Thanks Arav

$\endgroup$
1
$\begingroup$

I found Auto-encoders to be the best solution for this. Performing auto encoder before clustering reduces the dimensionality of the high dimensional data and then the encoder results can be extracted and used for whatever method we want to implement like,

supervised classification unsupervised clustering etc.

Autoencoders using CNN (encoder and decoder) for image: https://www.datacamp.com/community/tutorials/autoencoder-keras-tutorial#comment-6125

Autoencoders - its types and usage - LSTM: https://machinelearningmastery.com/lstm-autoencoders/

for my usecase I should be performing an embedding before feeding it into the autoencoder (CNN/LSTM) with only the encoder so that the reduced dimensions can be used for kmeans clustering.

Hope this helps people who have same question, I wonder how none has come across this situation or haven't replied for this.

Thanks Arav

$\endgroup$

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.