Tidyverse

From 太極
Revision as of 15:18, 23 August 2019 by Brb (talk | contribs) (→‎Tidyverse)
Jump to navigation Jump to search

Tidyverse

   Import
     |
     | readr, readxl
     | haven, DBI, httr   +----- Visualize ------+
     |                    |    ggplot2, ggvis    |
     |                    |                      |
   Tidy ------------- Transform 
   tibble               dplyr                   Model 
   tidyr                  |                    broom
                          +------ Model ---------+

Install on Ubuntu

How to install Tidyverse on Ubuntu 16.04 and 17.04

# Ubuntu >= 18.04. However, I get unmet dependencies errors on R 3.5.3.
# r-cran-curl : Depends: r-api-3.4
sudo apt-get install r-cran-curl r-cran-openssl r-cran-xml2

# Works fine on Ubuntu 16.04, 18.04
sudo apt install libcurl4-openssl-dev libssl-dev libxml2-dev

80 R packages will be installed after tidyverse has been installed.

Install on Raspberry Pi/(ARM based) Chromebook

In additional to the requirements of installing on Ubuntu, I got an error when it is installing a dependent package fs: undefined symbol: pthread_atfork. The fs package version is 1.2.6. The solution is to add one line in fs/src/Makevars file and then install the "fs" package using the source on the local machine.

5 most useful data manipulation functions

  • subset() for making subsets of data (natch)
  • merge() for combining data sets in a smart and easy way
  • melt()-reshape2 package for converting from wide to long data formats. See an example here where we want to combine multiple columns of values into 1 column. melt() is replaced by gather().
  • dcast()-reshape2 package for converting from long to wide data formats (or just use tapply()), and for making summary tables
  • ddply()-plyr package for doing split-apply-combine operations, which covers a huge swath of the most tricky data operations

data.table

Fast aggregation of large data (e.g. 100GB in RAM or just several GB size file), fast ordered joins, fast add/modify/delete of columns by group using no copies at all, list columns and a fast file reader (fread).

Some resources:

  • https://www.rdocumentation.org/packages/data.table/versions/1.12.0
  • R Packages: dplyr vs data.table
  • Cheat sheet from RStudio
  • Reading large data tables in R. fread(FILENAME)
  • Note that 'x[, 2] always return 2. If you want to do the thing you want, use x[, 2, with=FALSE] or x[, V2] where V2 is the header name. See the FAQ #1 in data.table.
  • Understanding data.table Rolling Joins
  • Intro to The data.table Package
    • Subsetting rows and/or columns
    • Alternative to using tapply(), aggregate(), table() to summarize data
    • Similarities to SQL, DT[i, j, by]
  • R : data.table (with 50 examples) from ListenData
    • Describe Data
    • Selecting or Keeping Columns
    • Rename Variables
    • Subsetting Rows / Filtering
    • Faster Data Manipulation with Indexing
    • Performance Comparison
    • Sorting Data
    • Adding Columns (Calculation on rows)
    • How to write Sub Queries (like SQL)
    • Summarize or Aggregate Columns
    • GROUP BY (Within Group Calculation)
    • Remove Duplicates
    • Extract values within a group
    • SQL's RANK OVER PARTITION
    • Cumulative SUM by GROUP
    • Lag and Lead
    • Between and LIKE Operator
    • Merging / Joins
    • Convert a data.table to data.frame
  • R Tutorial: data.table from dezyre.com
    • Syntax: DT[where, select|update|do, by]
    • Keys and setkey()
    • Fast grouping using j and by: DT[,sum(v),by=x]
    • Fast ordered joins: X[Y,roll=TRUE]
  • In the Introduction to data.table vignette, the data.table::order() function is SLOWER than base::order() from my Odroid xu4 (running Ubuntu 14.04.4 trusty on uSD)
    odt = data.table(col=sample(1e7))
    (t1 <- system.time(ans1 <- odt[base::order(col)]))  ## uses order from base R
    #   user  system elapsed 
    #  2.730   0.210   2.947 
    (t2 <- system.time(ans2 <- odt[order(col)]))        ## uses data.table's order
    #   user  system elapsed 
    #  2.830   0.215   3.052
    (identical(ans1, ans2))
    # [1] TRUE
  • Boost Your Data Munging with R
  • rbindlist(). One problem, it uses too much memory. In fact, when I try to analyze R package downloads, the command "dat <- rbindlist(logs)" uses up my 64GB memory (OS becomes unresponsive).

OpenMP enabled compiler for Mac. This instruction works on my Mac El Capitan (10.11.6) when I need to upgrade the data.table version from 1.11.4 to 1.11.6.

Question: how to make use multicore with data.table package?

reshape & reshape2 (superceded by tidyr package)

tidyr and benchmark

An evolution of reshape2. It's designed specifically for data tidying (not general reshaping or aggregating) and works well with dplyr data pipelines.

Make wide tables long with gather() (see 6.3.1 of Efficient R Programming)

library(tidyr)
library(efficient)
data(pew) # wide table
dim(pew) # 18 x 10,  (religion, '<$10k', '$10--20k', '$20--30k', ..., '>150k') 
pewt <- gather(data = pew, key = Income, value = Count, -religion)
dim(pew) # 162 x 3,  (religion, Income, Count)

args(gather)
# function(data, key, value, ..., na.rm = FALSE, convert = FALSE, factor_key = FALSE)

where the three arguments of gather() requires:

  • data: a data frame in which column names will become row vaues
  • key: the name of the categorical variable into which the column names in the original datasets are converted.
  • value: the name of cell value columns

In this example, the 'religion' column will not be included (-religion).

dplyr, plyr packages

  • plyr package suffered from being slow in some cases. dplyr addresses this by porting much of the computation to C++. Another additional feature is the ability to work with data stored directly in an external database. The benefits of doing this are the data can be managed natively in a relational database, queries can be conducted on that database, and only the results of query returned.
  • Essential functions: 3 rows functions, 3 column functions and 1 mixed function.
           select, mutate, rename
            +------------------+
filter      +                  +
arrange     +                  +
group_by    +                  +
drop_na     +                  +
ungroup     + summarise        +
            +------------------+
  • These functions works on data frames and tibble objects. Note stats package also has a filter() function for time series data. If we have not loaded the dplyr package, the filter() function below will give an error (count() also is from dplyr).
iris %>% filter(Species == "setosa") %>% count()
head(iris %>% filter(Species == "setosa") %>% arrange(Sepal.Length))
  • dplyr tutorial from PH525x series (Biomedical Data Science by Rafael Irizarry and Michael Love). For select() function, some additional options to select columns based on a specific criteria include
    • start_with()/ ends_with() = Select columns that start/end with a character string
    • contains() = Select columns that contain a character string
    • matches() = Select columns that match a regular expression
    • one_of() = Select columns names that are from a group of names
  • Data Transformation in the book R for Data Science. Five key functions in the dplyr package:
    • Filter rows: filter()
    • Arrange rows: arrange()
    • Select columns: select()
    • Add new variables: mutate()
    • Grouped summaries: group_by() & summarise()
# filter
jan1 <- filter(flights, month == 1, day == 1)
filter(flights, month == 11 | month == 12)
filter(flights, arr_delay <= 120, dep_delay <= 120)
df <- tibble(x = c(1, NA, 3))
filter(df, x > 1)
filter(df, is.na(x) | x > 1)

# arrange
arrange(flights, year, month, day)
arrange(flights, desc(arr_delay))

# select
select(flights, year, month, day)
select(flights, year:day)
select(flights, -(year:day))

# mutate
flights_sml <- select(flights, 
  year:day, 
  ends_with("delay"), 
  distance, 
  air_time
)
mutate(flights_sml,
  gain = arr_delay - dep_delay,
  speed = distance / air_time * 60
)
# if you only want to keep the new variables
transmute(flights,
  gain = arr_delay - dep_delay,
  hours = air_time / 60,
  gain_per_hour = gain / hours
)

# summarise()
by_day <- group_by(flights, year, month, day)
summarise(by_day, delay = mean(dep_delay, na.rm = TRUE))

# pipe. Note summarise() can return more than 1 variable.
delays <- flights %>% 
  group_by(dest) %>% 
  summarise(
    count = n(),
    dist = mean(distance, na.rm = TRUE),
    delay = mean(arr_delay, na.rm = TRUE)
  ) %>% 
  filter(count > 20, dest != "HNL")
flights %>% 
  group_by(year, month, day) %>% 
  summarise(mean = mean(dep_delay, na.rm = TRUE))

stringr

magrittr

x %>% f     # f(x)
x %>% f(y)  # f(x, y)
x %>% f(arg=y)  # f(x, arg=y)
x %>% f(z, .) # f(z, x)
x %>% f(y) %>% g(z)  #  g(f(x, y), z)

x %>% select(which(colSums(!is.na(.))>0))  # remove columns with all missing data
x %>% select(which(colSums(!is.na(.))>0)) %>% filter((rowSums(!is.na(.))>0)) # remove all-NA columns _and_ rows
suppressPackageStartupMessages(library("dplyr"))
starwars %>%
  filter(., height > 200) %>%
  select(., height, mass) %>%
  head(.)
# instead of 
starwars %>%
  filter(height > 200) %>%
  select(height, mass) %>%
  head
iris$Species
iris[["Species"]]

iris %>%
`[[`("Species")

iris %>%
`[[`(5)

iris %>%
  subset(select = "Species")
  • Split-apply-combine: group + summarize + sort/arrange + top n. The following example is from Efficient R programming.
data(wb_ineq, package = "efficient")
wb_ineq %>% 
  filter(grepl("g", Country)) %>%
  group_by(Year) %>%
  summarise(gini = mean(gini, na.rm  = TRUE)) %>%
  arrange(desc(gini)) %>%
  top_n(n = 5)
f <- function(x) {
  (y - x) %>% 
    '^'(2) %>% 
    sum %>%
    '/'(length(x)) %>% 
    sqrt %>% 
    round(2)
}
# Examples from R for Data Science-Import, Tidy, Transform, Visualize, and Model
diamonds <- ggplot2::diamonds
diamonds2 <- diamonds %>% dplyr::mutate(price_per_carat = price / carat)

pryr::object_size(diamonds)
pryr::object_size(diamonds2)
pryr::object_size(diamonds, diamonds2)

rnorm(100) %>% matrix(ncol = 2) %>% plot() %>% str()
rnorm(100) %>% matrix(ncol = 2) %T>% plot() %>% str() # 'tee' pipe
    # %T>% works like %>% except that it returns the lefthand side (rnorm(100) %>% matrix(ncol = 2))  
    # instead of the righthand side.

# If a function does not have a data frame based api, you can use %$%.
# It explodes out the variables in a data frame.
mtcars %$% cor(disp, mpg) 

# For assignment, magrittr provides the %<>% operator
mtcars <- mtcars %>% transform(cyl = cyl * 2) # can be simplified by
mtcars %<>% transform(cyl = cyl * 2)

Upsides of using magrittr: no need to create intermediate objects, code is easy to read.

When not to use the pipe

  • your pipes are longer than (say) 10 steps
  • you have multiple inputs or outputs
  • Functions that use the current environment: assign(), get(), load()
  • Functions that use lazy evaluation: tryCatch(), try()

outer()

Genomic sequence

  • chartr
> yourSeq <- "AAAACCCGGGTTTNNN"
> chartr("ACGT", "TGCA", yourSeq)
[1] "TTTTGGGCCCAAANNN"

broom

broom: Convert Statistical Analysis Objects into Tidy Tibbles

lobstr package - dig into the internal representation and structure of R objects

lobstr 1.0.0