New Extension: Interpretations - SHAP, LIME and Shapely

MartinLiebig

Dear Community

I am happy to announce that @pschlunder and I published a new extension to the marketplace: Interpretations!

So far RapidMiner users had the option to use Explain Predictions as their method to understand WHY a model predicted the way it did. The Explain Predictions operator uses an algorithm by @IngoRM which is focused around best speed, understand-ability and application on a range of use cases as well as data types.

The new operator adds the known algorithms of LIME, Shapely and SHAP to the mix. The operator Generate Interpretations has a very similar interface to the familiar Explain Predictions. In fact it also embeds Explain Predictions so that you can switch between different algorithms and get different ‘opinions’ on your predictions.

Please be aware that this is a first alpha release of the extension. We are continuously working on improving it. We appreciate every feedback!

 

Thanks!
Philipp & Martin

NeuralMarket

Very cool!

MartinLiebig

Small update on the extension:

Changes in 0.1.1
----------------
* Added an icon
* Fixed an error that there was no proper UserError if types of training and testing were different
* Fixed an error that real and integers were considered of different types.

Muhammed_Fatih_

Hi @mschmitz, 

thank you for the interesting insight and the integration of the 'Generate Interpretation' operator into RapidMiner! This can be very useful!

Is there a possibility to train and test a Decision Tree with Cross Validation by additionally using the KernelSHAP? I implemented the following procedure but it doesn't work due to the missing connections: 



Thank you in advance for your help! 

Best regards, 

Fatih  

MartinLiebig

Hi,

why would you do it inside? usually you use it afterwards with the final model.

Best,

Martin

anaRodrigues

Hi @mschmitz

I recently posted this question. It would be a big help if you could take a look.

Thank you in advance,
Ana

anaRodrigues

Hi @mschmitz,

Is it possible to use the generate interpretation operator with a model stored in the repository? I can't seem to make it work

Thanks,
Ana

MartinLiebig

that should work - what is the error message?

anms

Hi mschmitz,

I tried to use LIME, but I don't understand what does the value in circle means? Is there any guide to interpret the values in circle?

Thank you.

MartinLiebig

Hi @anms ,

LIME values are tricky to explain. Generally speaking: The higher the values, the more influential they are in the prediction of this specific example. Or maybe in other words: If you would change Age Category, the change of the confidence would be the biggest.

More - very technical - details can be found in this ebook - https://christophm.github.io/interpretable-ml-book/.

Best,

Martin

sectyn

What are the dependencies required to be installed for the Interpretation extension? Getting a missing dependency error after installing this extension.

MartinLiebig

Hi @sectyn ,

Only an up to date studio, i.e. 9.10.

Best,

Martin

hesu

After the upgrade to version 9.9, the previously arranged extensions cannot be opened, but can only enter the non-functional extension mode. How can I solve this problem

MartinLiebig

Hi hesu,

please check the rapidminer-studio.log for errors.

Best,

Martin

sherlock

Hi @mschmitz,

is there a way to connect it to the XGBoost Operator in Rapidminer? While e.g. GBM works fine, XGBoost doesn’t. Seems to be a type/interface issue. Taking into account the great performance of XGBoost it would be a pity if we couldn’t make use of this great operator for XGBoost.

Thanks
Erwin

MartinLiebig

Hi @sherlock ,

good point. I will check what XGBoost is implementing. I thought it works.

BR,

Martin