Recent developments in direct labeled graphicshttp://directlabels.r-forge.r-project.org

Toby Dylan Hockingtoby.hocking@inria.fr

29 March 2012

Motivation: confusing legends

How to add direct labels to some common plots

Recent developments in direct labeling

Conclusions

Problem 1: legend inconsistent with data

library(lattice)

dens <- densityplot(~score,loci,groups=type,

auto.key=list(space="top",columns=3),n=500,

main="Distribution of scores by selection type")

print(dens)

Distribution of scores by selection type

score

Density

0

2

4

6

8

10

0.0 0.5 1.0

Balancing Neutral Positive

Problem 2: too many classes render legend unreadable

data(BodyWeight,package="nlme")

library(ggplot2)

ratplot <- ggplot(BodyWeight,aes(Time,weight,colour=Rat))+

facet_grid(.~Diet)+

geom_line()

print(ratplot)

1 2 3

300

400

500

600

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60Time

weigh

t

Rat

2

3

4

1

8

5

6

7

11

9

10

12

13

15

14

16

Motivation: confusing legends

How to add direct labels to some common plots

Recent developments in direct labeling

Conclusions

The protocol I use for everyday plots in practice:

Do as many steps as needed until the plot is readable:

1. Make a lattice or ggplot2 plot p using colors and default legends.

2. Try the default direct labels: direct.label(p).

3. Check to see if another Positioning Method exists onhttp://directlabels.r-forge.r-project.org/docs/index.html

then use direct.label(p,"method").

4. If no Positioning Methods exist you can always write your own.

Add default direct labels at the mode of each density

library(directlabels)

direct.label(dens)

Distribution of scores by selection type

score

Density

0

2

4

6

8

10

0.0 0.5 1.0

Balancing

Neutral

Positive

With 2 groups, we label the min and max points

p <- xyplot(deaths~Time,uk.lung,

groups=sex,type=c("l","g"))

direct.label(p)

Time

deaths

500

1000

1500

2000

2500

1974 1975 1976 1977 1978 1979 1980

female

male

Label a scatterplot of the iris data by species

set.seed(1)

irisp <- xyplot(jitter(Sepal.Length)~jitter(Petal.Length),

iris,groups=Species)

direct.label(irisp)

jitter(Petal.Length)

jitter(Sepal.Length)

5

6

7

8

1 2 3 4 5 6 7

setosa

virginica

versicolor

Default direct labels for lineplots

direct.label(ratplot)

1 2 3

300

400

500

600

21385467

11910

12

13

161514

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60Time

weigh

t

Look up the Positioning Method on the directlabels website

direct.label(ratplot,"last.qp")

1 2 3

300

400

500

600

23418567

11910

12

13

151416

0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60Time

weigh

t

Construct your own custom Positioning Method

rp2 <- ratplot+

xlim(0,70)+ylim(150,650)

big.last <- list(cex=1.5,"last.qp")

direct.label(rp2,"big.last")

1 2 3

200

300

400

500

600

23418567

11910

12

13151416

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Direct label the LASSO path to visualize variables

lasso.plot <-

ggplot(path,aes(arclength,standardized.coef,colour=variable))+

geom_line(aes(group=variable))+

opts(title="LASSO path for prostate cancer data

calculated using the LARS")+

xlim(0,20)

direct.label(lasso.plot)

-2

0

2

4

6

lcp

age

gleason

lbphpgg45lweightsvi

lcavol

0 5 10 15 20

LASSO path for prostate cancer data calculated using the LARS

arclength

standardized.coef

Label the zero point to emphasize variable importance

direct.label(lasso.plot,"lasso.labels")

-2

0

2

4

6

lcavol

lweight

svi

lbph

pgg45

age

lcp

gleason

0 5 10 15 20

LASSO path for prostate cancer data calculated using the LARS

arclength

standardized.coef

Increase text size to make reading easier

direct.label(lasso.plot,list(cex=2,"lasso.labels"))

-2

0

2

4

6

lcavol

lweight

svi

lbph

pgg45

age

lcp

gleason

0 5 10 15 20

LASSO path for prostate cancer data calculated using the LARS

arclength

standardized.coef

Motivation: confusing legends

How to add direct labels to some common plots

Recent developments in direct labeling

Conclusions

Changes in recent versions of directlabels

directlabels version < 2.0 < 2.0 ≥ 2.0 ≥ 2.0plotting package lattice ggplot2 lattice ggplot2

basic Positioning Methods X X X Xsmart Positioning Methodsthat avoid label collisions X X X

redraw labelsafter window resize X X

fontface and fontfamily

text parameters X X Xlabel black and white plots X X X

label aestheticsother than color X

Label a scatterplot of the iris data by species

direct.label(irisp)

jitter(Petal.Length)

jitter(Sepal.Length)

5

6

7

8

1 2 3 4 5 6 7

setosa

virginica

versicolor

Show the grid for the searchFind a label position on a grid (grey rectangles)that is near the center of each point cloud (green dots),but does not overlap any points or other labels (black dots).

direct.label(irisp,debug=TRUE)

jitter(Petal.Length)

jitter(Sepal.Length)

5

6

7

8

1 2 3 4 5 6 7

setosa

virginica

versicolor

Increase text size before calculating label positions

rp2 <- ratplot+

xlim(0,70)+ylim(150,650)

big.last <- list(cex=1.5,"last.qp")

direct.label(rp2,"big.last")

1 2 3

200

300

400

500

600

23418567

11910

12

13151416

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Show the label borders used in the position calculation

direct.label(rp2,

list("big.last",

"calc.boxes",

"draw.rects"))

1 2 3

200

300

400

500

600

23418567

11910

12

13151416

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Label positions for lineplots are the solutions of a QP

t1

t2

tk−1

tk

...

h1

h2

hk−1

hk

...

Assume that for each text label i = 1, . . . , k we haveits position ti and height hi .Then optimal direct labels do not overlap, and areas close as possible to the target locations:

minb∈Rk

k∑i=1

(bi − ti )2 = ||b − t||2

subject to bi+1 ≥ bi + hi+1/2 + hi/2, ∀ i = 1, . . . , k − 1

This is a quadratic program (QP). QPs are convex sothere is a unique global minimum which correspondsto the best labels.We can solve this using quadprog::solve.QP()

and use the optimal b for the direct label positions.

Start with boxes at the end of each line

direct.label(rp2,list("last.points",cex=1.5,

"calc.boxes",

"draw.rects"))

1 2 3

200

300

400

500

600

12345678

1011

12

9

13141516

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Adjust box height if desired

direct.label(rp2,list("last.points",cex=1.5,"calc.boxes",

dl.trans(h=h+h/3),"calc.borders",

"draw.rects"))

1 2 3

200

300

400

500

600

12345678

1011

12

9

13141516

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Apply QP solver to get optimal labels

direct.label(rp2,list("last.points",cex=1.5,"calc.boxes",

dl.trans(h=h+h/3),"calc.borders",

qp.labels("y","h")))

1 2 3

200

300

400

500

600

23418567

11910

12

13151416

0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70Time

weigh

t

Motivation: confusing legends

How to add direct labels to some common plots

Recent developments in direct labeling

Conclusions

What directlabels is NOT

I Automatic publication-quality direct labels.(some manual tweaking will always be necessary)

I Optimal labels for individual points on scatterplots.(it is a bit more complicated)

-20 -10 0 10 20

-20

-10

010

20

x

y

-5.02192

1.31531

-0.78917

8.86785

1.16971

3.1863

-5.81791

7.14533

-8.25259

-3.59862

0.898860.96274

-2.01634

7.3984

1.2338

-0.29317

-3.88854

5.10856

-9.13814 23.10297

-4.3809

7.640612.61961

7.73405

-8.14379

-4.38451

-7.20222

2.30945

-11.57729

2.47076

-0.91114

17.57376

-1.3793

-1.11193

-6.90014

-2.21794

1.82908

4.17323

10.65402

9.70202

-1.01629

14.03203

-17.76776

6.22867

-5.22283

13.22231

-3.6344

13.19066

0.43779-18.78656

Use directlabels instead of confusing legends!

I Works with lattice and ggplot2.

I Sensible defaults.

I Useful in everyday plots in practice.

I Smart Positioning Methods that avoid label collisions.

I Customizable: you can write your own Positioning Methods.

Future work

I Automatically adjust xlim/ylim so labels stay on plot region?

I Contourplot labels as in contour()?

I Label using images instead of textual factor names?Possible Google Summer of Code 2012 project:http://rwiki.sciviews.org/doku.php?id=developers:projects:gsoc2012