I've previously used SHAP and LIME to explain predictions from a training set, i.e. I have the actual target value.

Is it possible to do the same to explain new predictions, i.e. I don't have the actual value to know how accurate my prediction is yet?

A use case could be forecasting a future value using a model trained on an existing time series. Whilst I can't tell how accurate the model is yet - it would be useful to know which features are contributing to a particular prediction.

  • 1
    $\begingroup$ Those kinds of post hoc methods actually need that you have confidence in your trained models. The explanations are for your model, not for the ground truth which is usually unknown. $\endgroup$
    – doubllle
    Commented Aug 5, 2020 at 11:46


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.