Gluon-ts: What is the best approach to explain a time series model from gluonts.

[MariPlaza]

We train a deep AR model for several time series using gluonts. Now we want to use libraries for explaining the model such as SHAP or LIME. Any suggestion on how to do it? Any other library that you know for doing it?

Thanks in advance

[StatMixedML]

Referring to https://github.com/awslabs/gluon-ts/issues/1013#issuecomment-687575025

[lostella]

@MariPlaza I am not familiar with SHAP or LIME, maybe you could explain what you have tried with those, or what blocks you from using them in the context of GluonTS models?

[MariPlaza]

@lostella SHAP and/or LIME are python packages to calculate Shapley values in order to explain the importance of the variables in black-box models. The topic is quite new and especially on multivariate time-series. We are at the beginning of our research for this project, and so far the best approach that I have read is the one proposed by @StatMixedML of using Temporal Fusion Transformer.

My question to you is if the TFT implementation on gluon-ts has (or will have) the interpretable insights as well as the PyTorch implementation has

Thanks to you and the whole team for this great package.

[lostella]

My question to you is if the TFT implementation on gluon-ts has (or will have) the interpretable insights as well as the PyTorch implementation has

@Gandor26 may help answering this

[Gandor26]

@lostella SHAP and/or LIME are python packages to calculate Shapley values in order to explain the importance of the variables in black-box models. The topic is quite new and especially on multivariate time-series. We are at the beginning of our research for this project, and so far the best approach that I have read is the one proposed by @StatMixedML of using Temporal Fusion Transformer.

My question to you is if the TFT implementation on gluon-ts has (or will have) the interpretable insights as well as the PyTorch implementation has

Thanks to you and the whole team for this great package.

@MariPlaza Currently we do not have an API to obtain the interpretable indicators, such as attention weights and variable importances, though they are computed within the model.
@lostella Does GluonTS estimators provide additional interfaces for such intermediate states? As far as I know, they only consider the final predictions for evaluation. Is there any workaround?