LIME and Sampling

[SebGGruber]

Still a total mess after the first chapters. Will reorder, merge, and delete lots of plots after some break with new ideas and after some criticism. The chapters below all the plots are not really planned chapters... just some kind of place with notes. Will probably kick out my proposal of a new sampling strategy and all related plots due to length reasons, but have to take with Christoph about this. Also wanted to show some mathematics related to this. Wanted to add a few more plots showing the superiority compared to default methods, but will most likely not have the space to do this. The rest of the missing text will be filled between the plots - expect no further chapters or topics.

[SebGGruber]

And please ignore the rmd file. 2100 lines of code.

[christophM]

• I agree with Philipp that the beginning of the chapter is already covered in the main LIME chapter and shouldn't be repeated here. You can assume that the reader knows LIME when reading your chapter. Only highlight parts of the theory that are very central to the sampling issue. For example to talk about the sampling step.
• Instead of barplots with error bars, I recommend plotting points for the mean coefficients with error bars. Similar to coefficient plots: http://repec.sowi.unibe.ch/stata/coefplot/getting-started.html . Better information to ink ratio and also bar plots are more associated with frequencies and not point estimates. Plots like FIGURE 2.19 will then look more informative.
• Put a paragraph in the beginning that outlines your chapter a bit. Here you should say what different scenarios you were looking at: numerical, categorical features, binning vs. no binning, ...

[christophM]

Your chapter has become a bit broader than only looking at what happens when you sample, because you also look at the binning. This is fine, but should be reflected in the beginning of the chapter. Just say something like that you are also looking at the effect of binning.