Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Better than My Meetup/Conference Talks: Going ...
Search
szilard
November 09, 2019
0
54
Better than My Meetup/Conference Talks: Going Deeper in Various GBM Topics - GBM Advanced Workshop - Budapest, Nov 2019
szilard
November 09, 2019
Tweet
Share
More Decks by szilard
See All by szilard
Gradient Boosting Machines (GBM): From Zero to Hero (with R and Python Code) - Data Con LA - Oct 2020
szilard
0
130
Make Machine Learning Boring Again: Best Practices for Using Machine Learning in Businesses - Albuquerque Machine Learning Meetup (Online) - Aug 2020
szilard
0
91
Better than Deep Learning: Gradient Boosting Machines (GBM) - eRum conference - invited talk - June 2020
szilard
0
90
Gradient Boosting Machines (GBM): From Zero to Hero (with R and Python Code) - LA Data Science Meetup - February 2020
szilard
0
83
A Random Walk in Data Science and Machine Learning in Practice - CEU, Business Analytics Masters - Budapest, Febr 2020
szilard
0
290
Gradient Boosting Machines (GBM): From Zero to Hero (with R and Python Code) - Budapest BI Forum, Budapest, Nov 2019
szilard
0
130
Make Machine Learning Boring Again: Best Practices for Using Machine Learning in Businesses - LA Data Science Meetup - Playa Vista, August 2019
szilard
0
100
Better than Deep Learning: Gradient Boosting Machines (GBM) / 2019 edition - Budapest R and Data Science Meetups - Budapest, June 2019
szilard
0
79
Better than Deep Learning: Gradient Boosting Machines (GBM) / 2019 edition - LA R Meetup - Santa Monica, May 2019
szilard
0
19
Featured
See All Featured
BBQ
matthewcrist
85
9.3k
RailsConf 2023
tenderlove
29
880
The Straight Up "How To Draw Better" Workshop
denniskardys
232
140k
Stop Working from a Prison Cell
hatefulcrawdad
267
20k
Docker and Python
trallard
40
3.1k
Mobile First: as difficult as doing things right
swwweet
222
8.9k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
43
6.6k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
26
2.1k
Bootstrapping a Software Product
garrettdimon
PRO
305
110k
Being A Developer After 40
akosma
86
590k
Imperfection Machines: The Place of Print at Facebook
scottboms
264
13k
Fireside Chat
paigeccino
32
3k
Transcript
Better than My Meetup/Conference Talks: Going Deeper in Various GBM
Topics Szilard Pafka, PhD Chief Scientist, Epoch (USA) GBM Advanced Workshop Budapest Nov 2019
Why GBMs
None
meetup/conference talks going deeper section dividers
None
Disclaimer: I am not representing my employer (Epoch) in this
talk I cannot confirm nor deny if Epoch is using any of the methods, tools, results etc. mentioned in this talk
Source: Andrew Ng
Source: Andrew Ng
Source: Andrew Ng
None
None
None
None
None
None
...
None
None
None
None
None
None
None
None
http://lowrank.net/nikos/pubs/empirical.pdf http://www.cs.cornell.edu/~alexn/papers/empirical.icml06.pdf
http://lowrank.net/nikos/pubs/empirical.pdf http://www.cs.cornell.edu/~alexn/papers/empirical.icml06.pdf
None
None
None
None
top algos (RF, boosting), all features 2007
top algos (RF, boosting), all features most algos (lin, tree,
nnet) worst algos (knn, NB) 2007
top algos (RF, boosting), all features most algos (lin, tree,
nnet) worst algos (knn, NB) top algos, removed top feature(s) 2007
None
None
None
Source: Hastie etal, ESL 2ed
Source: Hastie etal, ESL 2ed
GBM libs
None
None
None
None
10x
10x
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
Scoring
None
None
None
None
None
* very first request not shown >500ms (JVM “warmup”)
None
None
None
None
None
None
None
None
None
None
GBM-perf github repo
None
None
None
None
None
None
None
multi-core/socket
None
None
None
None
None
None
CPU 1
CPU 1 CPU 2
CPU 1 CPU 2
CPU 1 CPU 2
CPU 1 CPU 2
None
5x 3.5x
None
None
None
None
None
None
None
None
None
None
None
None
None
None
zero
None
None
Spark
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
None
GPU
None
None
None
None
catboost
None
None
None
None
None
None
None
None
None
None
API / tuning
None
None
None
None
None
http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf
http://www.argmin.net/2016/06/20/hypertuning/
http://www.argmin.net/2016/06/20/hypertuning/
None
None
None
time ordered data time ordered data
time ordered data time ordered data train sample
time ordered data time ordered data train test sample sample
(slightly different distribution)
time ordered data time ordered data train test sample sample
proper train early stopping Model selection resampled 80-10-10 (~CV) (slightly different distribution)
time ordered data time ordered data train test sample sample
proper train early stopping Model selection random search over lightgbm resampled 80-10-10 (~CV) (slightly different distribution)
None
None
None
None
None
None
None
None
Closing
None
None
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
Source: https://www.linkedin.com/pulse/winning-solution-kaggledays-2019-competition-san-francisco-mark-peng/
None
More:
None