Garson's algorithm TODO

Benjamin-Lee commented 4 years ago

In our black box rule, there is currently a TODO asking us to consider detailing Garson's algorithm.

I did some thinking and have the following thoughts:

Garson's algorithm hasn't fared well in empirical experiments: https://doi.org/10.1016/j.ecolmodel.2004.03.013 (pdf)
It's nearly impossible to find the original paper. I tried everything. The DOI doesn't point to anything (search on Google; the only hits are the ACM page and this paper) and I couldn't find it in the Harvard library or Sci-Hub.
I'm not sure we should spend the space in our paper describing algorithms in any detail, particularly ones whose usefulness in comparison to others is in question.

I suggest including a reference to a recent overview of DL interpretability for further reading. We could cite this paper (arXiv:2001.02522) and/or this freely available book (https://christophm.github.io/interpretable-ml-book/). I would prefer to cite the latter if at all possible, since it is also following the model of open-source collaborative writing that we are. Another advantage of the book Interpretable Machine Learning is that it contains links to code (usually in R and Python) for each of the methods it describes. I prefer to outsource the recommendation of actual software libraries to a living document (rather than a static paper such as this) so that it will retain its usefulness for longer.

I'll prep a PR with my suggestions in the coming days but would love to hear @tbrittoborges's thoughts on the matter.

rasbt commented 4 years ago

I suggest including a reference to a recent overview of DL interpretability for further reading. We could cite this paper (arXiv:2001.02522) and/or this freely available book (https://christophm.github.io/interpretable-ml-book/).

Why not recommending and referring to both?

tbrittoborges commented 4 years ago

I will follow this up later today.

Benjamin-Lee commented 4 years ago

I suggest including a reference to a recent overview of DL interpretability for further reading. We could cite this paper (arXiv:2001.02522) and/or this freely available book (https://christophm.github.io/interpretable-ml-book/).

Why not recommending and referring to both?

I'm happy to reference both. Working on a quick PR now.

tbrittoborges commented 4 years ago

I agree that both references are relevant, and that Fan et all is particular good as a tip. However, I would like discuss a few points on @Benjamin-Lee last PR #212

Benjamin-Lee commented 4 years ago

@tbrittoborges sounds good. I won't merge #212 until you comment on it.

Benjamin-Lee commented 4 years ago

@tbrittoborges I added some comments so let me know what you think.

Benjamin-Lee commented 4 years ago

Ok I rewrote the last sentence. Can I get some fresh reviews (maybe @tbrittoborges and/or @rasbt)?

Benjamin-Lee / deep-rules

Garson's algorithm TODO #209