Shap has this scatter plot functionality where you can plot on multiple instances in a scatter plot. On the x-axis one feature value is plotted. On the y-axis you can plot either (a) another feature value or (b) the shap score. Also you can color the instances again either by the shap scores or another feature. This is often used to see if some features interact with others or spot „islands of attributions“.
We could do the same and color the instances by the uncertainty predicted by the model. Either AU/EU or TU or some other measure.
This could be cool to see in what regions uncertainty is highest and spot anomalies.
Example:
Shap has this scatter plot functionality where you can plot on multiple instances in a scatter plot. On the x-axis one feature value is plotted. On the y-axis you can plot either (a) another feature value or (b) the shap score. Also you can color the instances again either by the shap scores or another feature. This is often used to see if some features interact with others or spot „islands of attributions“.
We could do the same and color the instances by the uncertainty predicted by the model. Either AU/EU or TU or some other measure.
This could be cool to see in what regions uncertainty is highest and spot anomalies.
Example: