Explainability, interpretability, and a half-successful trial

A refresher on definitions of these two terms:

Long story short, the task is to measure the explainability of a model. Traditionally, when it comes to explainability with tabular data, we can easily end up with something like SHapley Additive exPlanations.

How about an alternative model type? Specifically, a text-based classification model. shap actually has a pretty good adaptability in terms of text data. One of my favorite visualization output is the force plot from this example.

Another question is: How do we compare the explainability between two models with the same goal? In other words, how do you know model A makes more sense than model B?

If given the pre-condition that both models are trained and tested on the same data, we can probably calculate the proportion of top n-percentiles of tokens’ sum of absolute SHapley values over the sum of absolute SHapley values of all tokens.

The reasoning behind this is: for all tokens’ SHapley values, either positive or negative, is a relative contribution toward the final prediction. So the higher the value, the more decisive it is.

The above is just my immature proposal on how to compare models’ interpretability in terms of SHapley values.