Mark specific caveats

[jnothman]

eli5 tries to give general caveats about the interpretability of the weights for some particular class of models.

I wonder if there's a way to note more specific (if still somewhat naive) caveats, given a dataset, for example:

• feature is weighted very high but (for a linear model or similar) feature has very small range;
• feature is weighted only moderately high but there is another feature also weighted moderately high which is highly correlated in the dataset, i.e. the effective weight is understated;
• feature is weighted highly positive but there is another feature with a very negative weight which is highly correlated in the dataset, i.e. the effective weight is overstated.

The last couple of points may be indications that regularisation didn't work, for instance (e.g. grouped regularisation might be more appropriate).