Hadley Wickham @hadleywickham Chief Scientist, RStudio

Pipelines for data analysis in R

October 2015

Data analysis is the process by which data becomes

understanding, knowledge and insight

Data analysis is the process by which data becomes

understanding, knowledge and insight

Data analysis is the process by which data becomes

understanding, knowledge and insight

Data analysis is the process by which data becomes

understanding, knowledge and insight

Transform

Visualise

Model

Tidy

Import

Surprises, but doesn't scale

Scales, but doesn't (fundamentally) surprise

Create new variables & new summariesConsistent way of storing data

Transform

Visualise

Model

tidyr dplyrTidy

Importreadr readxl haven DBI httr

broom

ggplot2 ggvis

Pipelines

Think it Do itDescribe it

Cognitive

Computational

(precisely)

Cognition time ≫ Computation time

http://ww

w.flickr.com/photos/m

utsmuts/4695658106

%>%Inspirations: unix, F#, haskell, clojure, method chaining

magrittr::

foo_foo <- little_bunny()

bop_on( scoop_up( hop_through(foo_foo, forest), field_mouse ), head )

# vs foo_foo %>% hop_through(forest) %>% scoop_up(field_mouse) %>% bop_on(head)

x %>% f(y) # f(x, y)

x %>% f(z, .) # f(z, x)

x %>% f(y) %>% g(z) # g(f(x, y), z)

# Turns function composition (hard to read) # into sequence (easy to read)

# Any function can use it. Only needs a simple # property: the type of the first argument # needs to be the same as the type of the result.

# tidyr: pipelines for messy -> tidy data # dplyr: pipelines for data manipulation # ggvis: pipelines for visualisations # rvest: pipelines for html # purrr: pipelines for lists # xml2: pipelines for xml # stringr: pipelines for strings

Tidy

Transform

Visualise

Model

tidyr dplyrTidy

Importreadr readxl haven DBI httr

ggplot2 ggvis

broom

Storage Meaning

Table / File Data set

Rows Observations

Columns Variables

Tidy data = data that makes data analysis easy

Source: local data frame [5,769 x 22]

iso2 year m04 m514 m014 m1524 m2534 m3544 m4554 m5564 m65 mu f04 f514 (chr) (int) (int) (int) (int) (int) (int) (int) (int) (int) (int) (int) (int) (int) 1 AD 1989 NA NA NA NA NA NA NA NA NA NA NA NA 2 AD 1990 NA NA NA NA NA NA NA NA NA NA NA NA 3 AD 1991 NA NA NA NA NA NA NA NA NA NA NA NA 4 AD 1992 NA NA NA NA NA NA NA NA NA NA NA NA 5 AD 1993 NA NA NA NA NA NA NA NA NA NA NA NA 6 AD 1994 NA NA NA NA NA NA NA NA NA NA NA NA 7 AD 1996 NA NA 0 0 0 4 1 0 0 NA NA NA 8 AD 1997 NA NA 0 0 1 2 2 1 6 NA NA NA 9 AD 1998 NA NA 0 0 0 1 0 0 0 NA NA NA 10 AD 1999 NA NA 0 0 0 1 1 0 0 NA NA NA 11 AD 2000 NA NA 0 0 1 0 0 0 0 NA NA NA 12 AD 2001 NA NA 0 NA NA 2 1 NA NA NA NA NA 13 AD 2002 NA NA 0 0 0 1 0 0 0 NA NA NA 14 AD 2003 NA NA 0 0 0 1 2 0 0 NA NA NA 15 AD 2004 NA NA 0 0 0 1 1 0 0 NA NA NA 16 AD 2005 0 0 0 0 1 1 0 0 0 0 0 0 .. ... ... ... ... ... ... ... ... ... ... ... ... ... ... Variables not shown: f014 (int), f1524 (int), f2534 (int), f3544 (int), f4554 (int), f5564 (int), f65 (int), fu (int)

What are the variables in this dataset? (Hint: f = female, u = unknown, 1524 = 15-24)

# To convert this messy data into tidy data # we need two verbs. First we need to gather # together all the columns that aren't variables

tb2 <- tb %>% gather(demo, n, -iso2, -year, na.rm = TRUE) tb2

# Then separate the demographic variable into # sex and age tb3 <- tb2 %>% separate(demo, c("sex", "age"), 1) tb3

# Many tidyr verbs come in pairs: # spread vs. gather # extract/separate vs. unite # nest vs. unnest

Google for “tidyr” &

“tidy data”

Transform

Transform

Visualise

Model

tidyr dplyrTidy

ggplot2 ggvis

broom

Importreadr readxl haven DBI httr

Think it Do itDescribe it

Cognitive

Computational

(precisely)

One table verbs

• select: subset variables by name• filter: subset observations by value• mutate: add new variables• summarise: reduce to a single obs• arrange: re-order the observations

+ group by

Demo

right_join()

full_join()

inner_join()

left_join()

Mutating

semi_join()

anti_join()

Filtering Set

intersect()

setdiff()

union()

dplyr sources• Local data frame (C++)• Local data table• Local data cube (experimental)• RDMS: Postgres, MySQL, SQLite,

Oracle, MS SQL, JDBC, Impala• MonetDB, BigQuery

Google for “dplyr”

Visualise

Transform

Visualise

Model

tidyr dplyrTidy

ggplot2 ggvis

broom

Importreadr readxl haven DBI httr

What is ggvis?• A grammar of graphics

(like ggplot2)

• Reactive (interactive & dynamic) (like shiny)

• A pipeline (a la dplyr)

•Of the web (drawn with vega)

Demo4-ggvis.R 4-ggvis.Rmd

Google for “ggvis”

Modelwith broom, by David Robinson

Transform

Visualise

Model

tidyr dplyrTidy

ggplot2 ggvis

broom

Importreadr readxl haven DBI httr

2.5

5.0

7.5

1990 1995 2000 2005 2010 2015date

log(sales)

46 TX cities, ~25 years of data

What makes it hard to see the long term

trend?

# Models are useful as tool for removing # known patterns

tx <- tx %>% group_by(city) %>% mutate( resid = lm( log(sales) ~ factor(month), na.action = na.exclude ) %>% resid() )

−2

−1

0

1

1990 1995 2000 2005 2010 2015date

resid

# Models are also useful in their own right

models <- tx %>% group_by(city) %>% do(mod = lm( log(sales) ~ factor(month), data = ., na.action = na.exclude) )

Model summaries

• Model level: one row per model• Coefficient level: one row per

coefficient (per model)• Observation level: one row per

observation (per model)

Demo5-broom.R

Google for “broom r”

Big data and R

Big Can’t fit in memory on one computer: >5 TB

Medium Fits in memory on a server: 10 GB-5 TB

Small Fits in memory on a laptop: <10 GB

R is great at this!

R• R provides an excellent environment for

rapid interactive exploration of small data.

• There is no technical reason why it can’t also work well with medium size data. (But the work mostly hasn’t been done)

• What about big data?

1. Can be reduced to a small data problem with subsetting/sampling/summarising (90%)

2. Can be reduced to a very large number of small data problems (9%)

3. Is irreducibly big (1%)

The right small data

• Rapid iteration essential• dplyr supports this activity by avoiding

cognitive costs of switching between languages.

Lots of small problems

• Embarrassingly parallel (e.g. Hadoop)• R wrappers like foreach, rhipe, rhadoop• Challenging is matching architecture of

computing to data storage

Irreducibly big

• Computation must be performed by specialised system.

• Typically C/C++, Fortran, Scala.• R needs to be able to talk to those

systems.

Future work

End gameProvide a fluent interface where you spent your mental energy on the specific data problem, not general data analysis process.The best tools become invisible with time!Still a lot of work to do, especially on the connection between modelling and visualisation.

Transform

Visualise

Model

tidyr dplyrTidy

Importreadr readxl haven DBI httr

broom

ggplot2 ggvis