It's worth checking out Friedman's "Gradient Boosting Machine" paper (as mentioned here in the references) from 1999 -- this has a good description of "boosting" from the general perspective of function optimisation.
One interesting thing about Boosted Trees is the author's software (XGBoost[1]) reliably outperforms other implementations (in terms of accuracy of results[2]). I'm not entirely sure why this is - I know there is an open ticket in the Spark GBT implementation to investigate this.
I haven't read much about XGBoost boosted trees. Does each tree have additive independence? Is the tree ensemble of two trees better than one tree?
It seems like additive training that removes all constants in addition to regularization of model complexity would shape the tree ensemble into a baseline model that defines minimum assumptions. So, what's its success rate in predicting favorable outcomes vs. tree learning focused on heuristic specialization (impurity)?
In the first example, being male is one of two features that predict playing video games, and (surprise!) only the boy and the old man are classified as gamers. Talk about casual sexism! Can you imagine taking this class as a woman (who maybe, just maybe, happens to enjoy video games) and having to forgive/ignore the instructor's cluelessness in order to get through the material? So incredibly tone-deaf and lazy, ugh.
If it actually bothers you, then do something about it.
The author of these slides provided two versions of this material, one in prose (linked from my other comment in this thread) and these slides. Send a pull request to the XGBoost GitHub repo to update the prose version[1]. Once you're done with that, send an email to the author with the new images, non-abrasively explain why you're providing them, and politely ask the author to replace the slides on his website. Be sure to point out which slides need to be updated (8, 9, 24, 25, 28, 31) and which images go with which slides.
It's possible your email will get ignored (PhD students are busy and don't necessarily have time to maintain two year old teaching material), but I'm almost certain the pull request will be merged.
It actually does bother me, and I wish I had the time to fix it, but fixing it involves much more than the comically large amount of work you're asking of me: it means convincing this PhD student that he made a genuine mistake that betrays his sexism, and getting him to acknowledge that mistake by fixing his own slides.
I wish the instructor was more aware of his own sexism when he created this material, so that neither he nor I nor anyone would have to fix this problem now. He chose this careless example, and so it's his responsibility to fix it, no matter how busy he may be as a PhD student. We're all busy, but these slides have his name on them, and demonstrating that he knows better now is something only he can do.
Well, I could do it, change it to a more neutral example, then take a screenshot of your post and merge it to the repo. This would take maybe 2 hours (but I am familiar with git).
Fixing it would go very far to prove that you are sincere about the problem and isn't a hell of a lot of work. Going that far to fix it is more likely to convince him that it actually matters than anything else you could do.
10 comments
[ 3.4 ms ] story [ 36.8 ms ] threadHere's a copy: [pdf] http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.31....
[1] https://github.com/tqchen/xgboost
[2] It's also very fast in terms of absolute speed.
It seems like additive training that removes all constants in addition to regularization of model complexity would shape the tree ensemble into a baseline model that defines minimum assumptions. So, what's its success rate in predicting favorable outcomes vs. tree learning focused on heuristic specialization (impurity)?
https://www.fordfoundation.org/ideas/equals-change-blog/post...
http://www.pri.org/stories/2015-12-06/are-algorithms-racist-...
The author of these slides provided two versions of this material, one in prose (linked from my other comment in this thread) and these slides. Send a pull request to the XGBoost GitHub repo to update the prose version[1]. Once you're done with that, send an email to the author with the new images, non-abrasively explain why you're providing them, and politely ask the author to replace the slides on his website. Be sure to point out which slides need to be updated (8, 9, 24, 25, 28, 31) and which images go with which slides.
It's possible your email will get ignored (PhD students are busy and don't necessarily have time to maintain two year old teaching material), but I'm almost certain the pull request will be merged.
[1] https://github.com/dmlc/xgboost/blob/master/doc/model.md
PS: If you don't do this, I'll call you tone-deaf and lazy. ;)
I wish the instructor was more aware of his own sexism when he created this material, so that neither he nor I nor anyone would have to fix this problem now. He chose this careless example, and so it's his responsibility to fix it, no matter how busy he may be as a PhD student. We're all busy, but these slides have his name on them, and demonstrating that he knows better now is something only he can do.
Fixing it would go very far to prove that you are sincere about the problem and isn't a hell of a lot of work. Going that far to fix it is more likely to convince him that it actually matters than anything else you could do.