{treeheatr} and {pmlbr}

visualizing decision trees
on benchmark datasets

Trang Le

University of Pennsylvania

RLadies Miami 2020-11-19

@trang1618

Google doc: https://tiny.cc/treeheatr

Take a bad chart and make it better: a data visualization workshop

part 1

Your decision tree       may be cool, but what if I tell you
you can make it hot?

🌳

🔥

http://tiny.cc/treeheatr

decision tree + heatmap

http://tiny.cc/treeheatr

rpart.plot

Terry Therneau, Beth Atkinson, Brian Ripley

visNetwork::visTree()

Almende B.V., Benoit Thieurmel , Titouan Robert

partykit::plot.party()

Torsten Hothorn, Heidi Seibold, Achim Zeileis

ggparty

Martin Borkovec, Niyaz Madin, et al.

dtreeviz

(Python)

Terence Parr, Prince Grover

treeheatr

target

features

dependent variable/outcome/phenotype

predictors/variables

treeheatr::heat_tree(x)

can be of object

  • party (or constparty)
    specifying the precomputed tree
  • partynode 
    specifying the manual tree
  • data.frame (or tibble) 
heat_tree(x, target_lab = 'Outcome')

351 blood samples (January 10 - February 18)

3 features:

  • lactic dehydrogenase (LDH)
  • lymphocyte levels
  • high-sensitivity C-reactive protein (hs_CRP)

Yan et  al., 2020

COVID-19 patient data

http://tiny.cc/treeheatr

heat_tree(x = covid, target_lab = 'Outcome')

data.frame

heat_tree(covid, target_lab = 'Outcome',
	  feats = NA)
split_ldh <- partysplit(1L, breaks = 365)
split_crp <- partysplit(2L, breaks = 41.2)
split_lymp <- partysplit(3L, breaks = 14.7)

custom_tree <- partynode(1L, split = split_ldh , kids = list(
  partynode(2L, split = split_crp, kids = list(
    partynode(3L, info = 'Survival'),
    partynode(4L, split = split_lymp, kids = list(
      partynode(5L, info = 'Death'),
      partynode(6L, info = 'Survival'))))),
  partynode(7L, info = 'Death')))

heat_tree(x = custom_tree, data_test = covid, target_lab = 'Outcome')

http://tiny.cc/treeheatr

part 2

what is benchmarking?

a standard practice to illustrate the strengths and weaknesses of algorithms with regards to different problem characteristics

target

features

dependent variable/outcome/phenotype

predictors/variables

pmlbr::fetch_data(x)

character object

name of the dataset to fetch from PMLB

fetch_data('wine_quality_red')
fixed.acidity volatile.acidity ... target
7.4 0.700 5
7.8 0.880 5
7.8 0.760 5
11.2 0.280 6
7.4 0.700 5
7.4 0.660 5
7.9 0.600 5 
fetch_data('wine_quality_red')

PMLB for other purposes

teaching

first time open source contributors

http://tiny.cc/treeheatr

Thanks!

@trang1618

Funding

NIH LM010098

NIH AI116794

R packages

ggplot2

partykit

ggparty

heatmaply

People

Jason Moore

Anonymous reviewers

http://tiny.cc/treeheatr

Made with Slides.com