This ticket address some tuning that could be done in LIME before putting it in use, examples of such optimizations are:

• Use bootstrapping from statistics to generate feature intervals for more precise perturbation for numerical values (at least) from a few number of inputs.
• Use existing/past predictions / saliencies to generate stability metrics and fit LIME parameters to reach best stability 
• Use same procedure above but with impact-score alone or in combination with other metrics...