incremental CART decision tree, based on the hoeffding tree i.e. very fast decision tree (VFDT), which is proposed in this paper "Mining High-Speed Data Streams" by Domingos & Hulten (2000). And a newly extended model "Extremely Fast Decision Tree" (EFDT) by Manapragada, Webb & Salehi (2018). Added new implementation of Random Forest
100
stars
28
forks
source link
update and predict now take sklearn-style arrays (+ PEP8) #2
I've been using your VFDT code, thank you a lot, it's very clear and well-made.
I'm using VFDT in a pipeline where I sometimes want to pass it arrays to update/predict at once, so:
I renamed your (VFDT class') update & predict (that take single vectors) into __update and __predict and created update & predict to call check_array and check_X_y from sklearn.utils before calling your own functions.
I also wanted to try the algorithm without tie-breaking, so I added a condition to disable tie-breaking when tau == 0. It was a terrible idea, the algo does not converge, but still, it's a feature.
Finally, I use a formatter/linter that is a PEP8 nazi on my text editor, so it cut every line that was > 80 chars into several lines. This is just cosmetic.
(Your example at the bottom of vfdt.py still works.)
(I didn't open the other files.)
In case you are interested, have this pull request, with regards,
Adrien Luxey
PS: If you want to be even more sklearn compatible, you should just rename your update into fit. You could also add a fit_predict function doing both, as in e.g. KMeans.
Hi!
I've been using your VFDT code, thank you a lot, it's very clear and well-made.
I'm using VFDT in a pipeline where I sometimes want to pass it arrays to update/predict at once, so:
update&predict(that take single vectors) into__updateand__predictand createdupdate&predictto callcheck_arrayandcheck_X_yfromsklearn.utilsbefore calling your own functions.tau == 0. It was a terrible idea, the algo does not converge, but still, it's a feature.vfdt.pystill works.)In case you are interested, have this pull request, with regards, Adrien Luxey
PS: If you want to be even more sklearn compatible, you should just rename your
updateintofit. You could also add afit_predictfunction doing both, as in e.g. KMeans.