Create SAX-based Shap explainer (Univariate time-series)

[wei-enwang]

This new explainer will be composed of three modules. The first module discovers the motifs existing in the dataset and produces the set of motifs. The second module analyzes the instance signal to be explained by the explainer by finding the location of the motifs inside the signal. With the results of the second module, the last module featurizes the signal with those motifs, which allows us to "turn on/off" the motifs in the signal as means of perturbation/occlusion to help SHAP calculations.

The first module will be based on HIME, a SAX-based motif discovery algorithm. The second part could be some pattern matching algorithm or an extension of the preceding module. The final part requires the transformation from the "motif space" back to the raw signal space followed by a standard SHAP explainer(KernelSHAP or the default version).