What’s the difference between performing counterfactual explanation on tabular data versus time-series data? Can we apply the same strategies to time-series data, and are the properties that generate a good counterfactual also valid for time-series data counterfactual explanation?
When dealing with time-series data, the main difference is that we try to find a continuous subsequence that has the greatest impact on the classification [@guidottiExplainingAnyTime2020][@delaneyInstancebasedCounterfactualExplanations2021], in order to modify it and change the classification. This is different from non-sequential data, where there is no notion of continuous subsequence.
As stated by [@delaneyInstancebasedCounterfactualExplanations2021], it’s not clear if the counterfactual explanation for tabular data and other types of data can be transferred for time-series data, and if the properties a good counterfactual are actually satisfied.
On the other hand, in the paper by [@schlegelRigorousEvaluationXAI2019], they show how some explanation methods tailored for some types of data (like images and text) such as LIME [@ribeiroWhyShouldTrust2016] and SHAP [@lundbergUnifiedApproachInterpreting] actually works well also with time-series data.
In the paper by [@guidottiExplainingAnyTime2020], they address the procedure done by Schlegel et al., saying how the data is pre-processed with a-priori segmentation of the time series, which can cause a loss of information. Affirming also that the advantage of a explainability method that is tailored for time-series data (like the one proposed by Guidotti et al.) is the non-need to do any a-priori segmentation or other types of pre-processing to the data.