Try to incorporate SHAP to aid model interpretability

[yiwen-h]

https://shap.readthedocs.io/en/latest/index.html

https://medium.com/nlplanet/two-minutes-nlp-explain-predictions-with-shap-values-2a0e34219177