Can anyone provide an example workflow using the Compute LIME component?

KNIME Analytics Platform
1

Hi,

I am trying to interpret the predicted probability for each record for my analysis. Can anyone help me in providing an example workflow with compute LIME component? I am trying to learn what it does.

2

Hi @Siddartha4sights and welcome to the forum. Check out this workflow from Paolo:

https://kni.me/w/3NciU4lnW6e4RMk1

(BTW, great meeting you at Summit last week!)

2 Likes
3

Thanks Scott. I used the workflow as a base and created my own workflow. Could you explain the output values of the LIME component?

4

You can check the description pane of the Compute LIME component to see what it’s generating (you’ll have to scroll down a bit to see what the output port contains). But since it’s used in conjunction with the LIME Loop Start node you really need to look at both.

The idea here is that since it can be hard to interpret the results of an ensemble model like XGBoost, maybe we can train a linear model local to the feature space of a particular prediction, and look at the influence at a particular spot, which should be a little easier to understand.

Basically, if you back up to the LIME Loop Start, a few particular wines with either high or low sulfates are chosen for explanation. For each for these 18 wines, a local model is trained within the loop. The coefficients of these individual local models are generated by the Compute LIME component, which are in turn collected by the Loop End node. We can examine these coefficients to try to get an idea of which features in the model have the most influence on the target variable, which is what the visualization node at the end does.

For example, high sulphate wines with true positive classifications are most strongly influenced by chlorides, volatile acidity, and sulphates.

Does that help?

5

This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.