Closed wang99711123 closed 1 year ago
Yes, but the support of time series forecasting in the current version is limited. It only includes SHAP for time series forecasting, and the running time may be long if the model uses a long historical data for forecasting. We will improve this part in the future release.
I would like to know, how to use the interpretability of time series forecasting in the current version? thank you!
It is "feature-attribution" explanation. For example, the forecasting model is xt = f(x{t-1}, x{t-2}, ..., x{t-k}). It will compute the feature attribution/importance scores of x{t-1}, x{t-2}, ..., x_{t-k} when the model predicts xt. The lib also computes the feature importance scores for the timestamps besides the observed values x{t-1}, x{t-2}, ..., x{t-k}.
OK I see, thanks a lot!
Thanks for the excellent work! And I wonder if OmniXAI supports interpretability of time series forecasting?