-Marcus

]]>I think I will first try this Lek method on my keras model and then the mentioned below.

If you are interested:

Most approaches to interpret the feature Importance I found are based on the visualization of the network, activation functions etc.

However in my research, I discovered three methods for RNN’s/LSTM’s.

Gradient-based Sensitivity Analysis and LRP:

https://arxiv.org/pdf/1706.07206.pdf

Garson’s method:

https://www.biorxiv.org/content/biorxiv/early/2016/09/28/078246.full.pdf

Greetings, Chris

]]>*******

Hello,

I really like your post’s about the sensitivity analysis and the variable importance. Have you already used the sensitivity analysis (any method of it) on RNN or LSTM Networks? What do you think about this? Generally this should work or am I wrong?

Thanks! ]]>

The sensitivity analysis (Lek Profile Method) should work with any R model object that has a predict method. That function relies only on the fitted model object, so the structure of the weights is not important. The variable importance functions are a bit different as they extract the weights from a fitted model object to estimate importance. I am not as familiar with RNN or LSTM, but my guess is the methods are not as easily transferable. The variable importance functions were designed for simple feed-forward networks and the weights for the other model types likely do not have the same form. Check the source code here:https://github.com/fawda123/NeuralNetTools/tree/master/R You may find some useful snippets that you can adapt for these other models.

-Marcus

]]>WordPress has some markup tags you can use, with an option to specify the coding language: https://en.support.wordpress.com/code/posting-source-code/

-Marcus

]]>I really like your post’s about the sensitivity analysis and the variable importance. Have you already used the sensitivity analysis (any method of it) on RNN or LSTM Networks? What do you think about this? Generally this should work or am I wrong?

Thanks! ]]>

If it is possible then please suggest me.

Many thanks

Javid

-Marcus

]]>