SparkR - Practical Guide
sparkr-vignettes.Rmd
Overview
SparkR is an R package that provides a light-weight frontend to use Apache Spark from R. With Spark 3.4.1, SparkR provides a distributed data frame implementation that supports data processing operations like selection, filtering, aggregation etc. and distributed machine learning using MLlib.
Getting Started
We begin with an example running on the local machine and provide an overview of the use of SparkR: data ingestion, data processing and machine learning.
First, let’s load and attach the package.
SparkSession
is the entry point into SparkR which connects your R program to a Spark cluster. You can create a SparkSession
using sparkR.session
and pass in options such as the application name, any Spark packages depended on, etc.
We use default settings in which it runs in local mode. It auto downloads Spark package in the background if no previous installation is found. For more details about setup, see Spark Session.
## Java ref type org.apache.spark.sql.SparkSession id 1
The operations in SparkR are centered around an R class called SparkDataFrame
. It is a distributed collection of data organized into named columns, which is conceptually equivalent to a table in a relational database or a data frame in R, but with richer optimizations under the hood.
SparkDataFrame
can be constructed from a wide array of sources such as: structured data files, tables in Hive, external databases, or existing local R data frames. For example, we create a SparkDataFrame
from a local R data frame,
cars <- cbind(model = rownames(mtcars), mtcars)
carsDF <- createDataFrame(cars)
We can view the first few rows of the SparkDataFrame
by head
or showDF
function.
head(carsDF)
## model mpg cyl disp hp drat wt qsec vs am gear carb
## 1 Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4
## 2 Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4
## 3 Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1
## 4 Hornet 4 Drive 21.4 6 258 110 3.08 3.215 19.44 1 0 3 1
## 5 Hornet Sportabout 18.7 8 360 175 3.15 3.440 17.02 0 0 3 2
## 6 Valiant 18.1 6 225 105 2.76 3.460 20.22 1 0 3 1
Common data processing operations such as filter
and select
are supported on the SparkDataFrame
.
carsSubDF <- select(carsDF, "model", "mpg", "hp")
carsSubDF <- filter(carsSubDF, carsSubDF$hp >= 200)
head(carsSubDF)
## model mpg hp
## 1 Duster 360 14.3 245
## 2 Cadillac Fleetwood 10.4 205
## 3 Lincoln Continental 10.4 215
## 4 Chrysler Imperial 14.7 230
## 5 Camaro Z28 13.3 245
## 6 Ford Pantera L 15.8 264
SparkR can use many common aggregation functions after grouping.
## gear count
## 1 4 12
## 2 3 15
## 3 5 5
The results carsDF
and carsSubDF
are SparkDataFrame
objects. To convert back to R data.frame
, we can use collect
. Caution: This can cause your interactive environment to run out of memory, though, because collect()
fetches the entire distributed DataFrame
to your client, which is acting as a Spark driver.
## [1] "data.frame"
SparkR supports a number of commonly used machine learning algorithms. Under the hood, SparkR uses MLlib to train the model. Users can call summary
to print a summary of the fitted model, predict
to make predictions on new data, and write.ml
/read.ml
to save/load fitted models.
SparkR supports a subset of R formula operators for model fitting, including ‘~’, ‘.’, ‘:’, ‘+’, and ‘-‘. We use linear regression as an example.
model <- spark.glm(carsDF, mpg ~ wt + cyl)
The result matches that returned by R glm
function applied to the corresponding data.frame
mtcars
of carsDF
. In fact, for Generalized Linear Model, we specifically expose glm
for SparkDataFrame
as well so that the above is equivalent to model <- glm(mpg ~ wt + cyl, data = carsDF)
.
summary(model)
##
## Deviance Residuals:
## (Note: These are approximate quantiles with relative error <= 0.01)
## Min 1Q Median 3Q Max
## -4.2893 -1.7085 -0.4713 1.5729 6.1004
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 39.6863 1.71498 23.1409 0.00000000
## wt -3.1910 0.75691 -4.2158 0.00022202
## cyl -1.5078 0.41469 -3.6360 0.00106428
##
## (Dispersion parameter for gaussian family taken to be 6.592137)
##
## Null deviance: 1126.05 on 31 degrees of freedom
## Residual deviance: 191.17 on 29 degrees of freedom
## AIC: 156
##
## Number of Fisher Scoring iterations: 1
The model can be saved by write.ml
and loaded back using read.ml
.
write.ml(model, path = "/HOME/tmp/mlModel/glmModel")
In the end, we can stop Spark Session by running
Setup
Installation
Different from many other R packages, to use SparkR, you need an additional installation of Apache Spark. The Spark installation will be used to run a backend process that will compile and execute SparkR programs.
After installing the SparkR package, you can call sparkR.session
as explained in the previous section to start and it will check for the Spark installation. If you are working with SparkR from an interactive shell (e.g. R, RStudio) then Spark is downloaded and cached automatically if it is not found. Alternatively, we provide an easy-to-use function install.spark
for running this manually. If you don’t have Spark installed on the computer, you may download it from Apache Spark Website.
If you already have Spark installed, you don’t have to install again and can pass the sparkHome
argument to sparkR.session
to let SparkR know where the existing Spark installation is.
sparkR.session(sparkHome = "/HOME/spark")
Spark Session
In addition to sparkHome
, many other options can be specified in sparkR.session
. For a complete list, see Starting up: SparkSession and SparkR API doc.
In particular, the following Spark driver properties can be set in sparkConfig
.
Property Name | Property group | spark-submit equivalent |
---|---|---|
spark.driver.memory |
Application Properties | --driver-memory |
spark.driver.extraClassPath |
Runtime Environment | --driver-class-path |
spark.driver.extraJavaOptions |
Runtime Environment | --driver-java-options |
spark.driver.extraLibraryPath |
Runtime Environment | --driver-library-path |
spark.kerberos.keytab |
Application Properties | --keytab |
spark.kerberos.principal |
Application Properties | --principal |
For Windows users: Due to different file prefixes across operating systems, to avoid the issue of potential wrong prefix, a current workaround is to specify spark.sql.warehouse.dir
when starting the SparkSession
.
spark_warehouse_path <- file.path(path.expand('~'), "spark-warehouse")
sparkR.session(spark.sql.warehouse.dir = spark_warehouse_path)
Cluster Mode
SparkR can connect to remote Spark clusters. Cluster Mode Overview is a good introduction to different Spark cluster modes.
When connecting SparkR to a remote Spark cluster, make sure that the Spark version and Hadoop version on the machine match the corresponding versions on the cluster. Current SparkR package is compatible with
## [1] "Spark 3.4.1"
It should be used both on the local computer and on the remote cluster.
To connect, pass the URL of the master node to sparkR.session
. A complete list can be seen in Spark Master URLs. For example, to connect to a local standalone Spark master, we can call
sparkR.session(master = "spark://local:7077")
For YARN cluster, SparkR supports the client mode with the master set as “yarn”.
sparkR.session(master = "yarn")
Yarn cluster mode is not supported in the current version.
Data Import
Local Data Frame
The simplest way is to convert a local R data frame into a SparkDataFrame
. Specifically we can use as.DataFrame
or createDataFrame
and pass in the local R data frame to create a SparkDataFrame
. As an example, the following creates a SparkDataFrame
based using the faithful
dataset from R.
df <- as.DataFrame(faithful)
head(df)
## eruptions waiting
## 1 3.600 79
## 2 1.800 54
## 3 3.333 74
## 4 2.283 62
## 5 4.533 85
## 6 2.883 55
Data Sources
SparkR supports operating on a variety of data sources through the SparkDataFrame
interface. You can check the Spark SQL Programming Guide for more specific options that are available for the built-in data sources.
The general method for creating SparkDataFrame
from data sources is read.df
. This method takes in the path for the file to load and the type of data source, and the currently active Spark Session will be used automatically. SparkR supports reading CSV, JSON and Parquet files natively and through Spark Packages you can find data source connectors for popular file formats like Avro. These packages can be added with sparkPackages
parameter when initializing SparkSession using sparkR.session
.
sparkR.session(sparkPackages = "com.databricks:spark-avro_2.12:3.0.0")
We can see how to use data sources using an example CSV input file. For more information please refer to SparkR read.df API documentation.
df <- read.df(csvPath, "csv", header = "true", inferSchema = "true", na.strings = "NA")
The data sources API natively supports JSON formatted input files. Note that the file that is used here is not a typical JSON file. Each line in the file must contain a separate, self-contained valid JSON object. As a consequence, a regular multi-line JSON file will most often fail.
Let’s take a look at the first two lines of the raw JSON file used here.
filePath <- paste0(sparkR.conf("spark.home"),
"/examples/src/main/resources/people.json")
readLines(filePath, n = 2L)
## [1] "{\"name\":\"Michael\"}" "{\"name\":\"Andy\", \"age\":30}"
We use read.df
to read that into a SparkDataFrame
.
## [1] 3
head(people)
## age name
## 1 NA Michael
## 2 30 Andy
## 3 19 Justin
SparkR automatically infers the schema from the JSON file.
printSchema(people)
## root
## |-- age: long (nullable = true)
## |-- name: string (nullable = true)
If we want to read multiple JSON files, read.json
can be used.
people <- read.json(paste0(Sys.getenv("SPARK_HOME"),
c("/examples/src/main/resources/people.json",
"/examples/src/main/resources/people.json")))
count(people)
## [1] 6
The data sources API can also be used to save out SparkDataFrames
into multiple file formats. For example we can save the SparkDataFrame
from the previous example to a Parquet file using write.df
.
write.df(people, path = "people.parquet", source = "parquet", mode = "overwrite")
Hive Tables
You can also create SparkDataFrames from Hive tables. To do this we will need to create a SparkSession with Hive support which can access tables in the Hive MetaStore. Note that Spark should have been built with Hive support and more details can be found in the SQL Programming Guide. In SparkR, by default it will attempt to create a SparkSession with Hive support enabled (enableHiveSupport = TRUE
).
sql("CREATE TABLE IF NOT EXISTS src (key INT, value STRING)")
txtPath <- paste0(sparkR.conf("spark.home"), "/examples/src/main/resources/kv1.txt")
sqlCMD <- sprintf("LOAD DATA LOCAL INPATH '%s' INTO TABLE src", txtPath)
sql(sqlCMD)
results <- sql("FROM src SELECT key, value")
# results is now a SparkDataFrame
head(results)
Data Processing
To dplyr users: SparkR has similar interface as dplyr in data processing. However, some noticeable differences are worth mentioning in the first place. We use df
to represent a SparkDataFrame
and col
to represent the name of column here.
indicate columns. SparkR uses either a character string of the column name or a Column object constructed with
$
to indicate a column. For example, to selectcol
indf
, we can writeselect(df, "col")
orselect(df, df$col)
.describe conditions. In SparkR, the Column object representation can be inserted into the condition directly, or we can use a character string to describe the condition, without referring to the
SparkDataFrame
used. For example, to select rows with value > 1, we can writefilter(df, df$col > 1)
orfilter(df, "col > 1")
.
Here are more concrete examples.
dplyr | SparkR |
---|---|
select(mtcars, mpg, hp) |
select(carsDF, "mpg", "hp") |
filter(mtcars, mpg > 20, hp > 100) |
filter(carsDF, carsDF$mpg > 20, carsDF$hp > 100) |
Other differences will be mentioned in the specific methods.
We use the SparkDataFrame
carsDF
created above. We can get basic information about the SparkDataFrame
.
carsDF
## SparkDataFrame[model:string, mpg:double, cyl:double, disp:double, hp:double, drat:double, wt:double, qsec:double, vs:double, am:double, gear:double, carb:double]
Print out the schema in tree format.
printSchema(carsDF)
## root
## |-- model: string (nullable = true)
## |-- mpg: double (nullable = true)
## |-- cyl: double (nullable = true)
## |-- disp: double (nullable = true)
## |-- hp: double (nullable = true)
## |-- drat: double (nullable = true)
## |-- wt: double (nullable = true)
## |-- qsec: double (nullable = true)
## |-- vs: double (nullable = true)
## |-- am: double (nullable = true)
## |-- gear: double (nullable = true)
## |-- carb: double (nullable = true)
SparkDataFrame Operations
Selecting rows, columns
SparkDataFrames support a number of functions to do structured data processing. Here we include some basic examples and a complete list can be found in the API docs:
You can also pass in column name as strings.
## mpg
## 1 21.0
## 2 21.0
## 3 22.8
## 4 21.4
## 5 18.7
## 6 18.1
Filter the SparkDataFrame to only retain rows with mpg less than 20 miles/gallon.
## model mpg cyl disp hp drat wt qsec vs am gear carb
## 1 Hornet Sportabout 18.7 8 360.0 175 3.15 3.44 17.02 0 0 3 2
## 2 Valiant 18.1 6 225.0 105 2.76 3.46 20.22 1 0 3 1
## 3 Duster 360 14.3 8 360.0 245 3.21 3.57 15.84 0 0 3 4
## 4 Merc 280 19.2 6 167.6 123 3.92 3.44 18.30 1 0 4 4
## 5 Merc 280C 17.8 6 167.6 123 3.92 3.44 18.90 1 0 4 4
## 6 Merc 450SE 16.4 8 275.8 180 3.07 4.07 17.40 0 0 3 3
Grouping, Aggregation
A common flow of grouping and aggregation is
Use
groupBy
orgroup_by
with respect to some grouping variables to create aGroupedData
objectFeed the
GroupedData
object toagg
orsummarize
functions, with some provided aggregation functions to compute a number within each group.
A number of widely used functions are supported to aggregate data after grouping, including avg
, count_distinct
, count
, first
, kurtosis
, last
, max
, mean
, min
, sd
, skewness
, stddev_pop
, stddev_samp
, sum_distinct
, sum
, var_pop
, var_samp
, var
. See the API doc for aggregate functions linked there.
For example we can compute a histogram of the number of cylinders in the mtcars
dataset as shown below.
## cyl count
## 1 8 14
## 2 4 11
## 3 6 7
Use cube
or rollup
to compute subtotals across multiple dimensions.
## SparkDataFrame[cyl:double, gear:double, am:double, avg(mpg):double]
generates groupings for {(cyl
, gear
, am
), (cyl
, gear
), (cyl
), ()}, while
## SparkDataFrame[cyl:double, gear:double, am:double, avg(mpg):double]
generates groupings for all possible combinations of grouping columns.
Operating on Columns
SparkR also provides a number of functions that can directly applied to columns for data processing and during aggregation. The example below shows the use of basic arithmetic functions.
carsDF_km <- carsDF
carsDF_km$kmpg <- carsDF_km$mpg * 1.61
head(select(carsDF_km, "model", "mpg", "kmpg"))
## model mpg kmpg
## 1 Mazda RX4 21.0 33.810
## 2 Mazda RX4 Wag 21.0 33.810
## 3 Datsun 710 22.8 36.708
## 4 Hornet 4 Drive 21.4 34.454
## 5 Hornet Sportabout 18.7 30.107
## 6 Valiant 18.1 29.141
Window Functions
A window function is a variation of aggregation function. In simple words,
aggregation function:
n
to1
mapping - returns a single value for a group of entries. Examples includesum
,count
,max
.window function:
n
ton
mapping - returns one value for each entry in the group, but the value may depend on all the entries of the group. Examples includerank
,lead
,lag
.
Formally, the group mentioned above is called the frame. Every input row can have a unique frame associated with it and the output of the window function on that row is based on the rows confined in that frame.
Window functions are often used in conjunction with the following functions: windowPartitionBy
, windowOrderBy
, partitionBy
, orderBy
, over
. To illustrate this we next look at an example.
We still use the mtcars
dataset. The corresponding SparkDataFrame
is carsDF
. Suppose for each number of cylinders, we want to calculate the rank of each car in mpg
within the group.
carsSubDF <- select(carsDF, "model", "mpg", "cyl")
ws <- orderBy(windowPartitionBy("cyl"), "mpg")
carsRank <- withColumn(carsSubDF, "rank", over(rank(), ws))
head(carsRank, n = 20L)
## model mpg cyl rank
## 1 Volvo 142E 21.4 4 1
## 2 Toyota Corona 21.5 4 2
## 3 Datsun 710 22.8 4 3
## 4 Merc 230 22.8 4 3
## 5 Merc 240D 24.4 4 5
## 6 Porsche 914-2 26.0 4 6
## 7 Fiat X1-9 27.3 4 7
## 8 Honda Civic 30.4 4 8
## 9 Lotus Europa 30.4 4 8
## 10 Fiat 128 32.4 4 10
## 11 Toyota Corolla 33.9 4 11
## 12 Merc 280C 17.8 6 1
## 13 Valiant 18.1 6 2
## 14 Merc 280 19.2 6 3
## 15 Ferrari Dino 19.7 6 4
## 16 Mazda RX4 21.0 6 5
## 17 Mazda RX4 Wag 21.0 6 5
## 18 Hornet 4 Drive 21.4 6 7
## 19 Cadillac Fleetwood 10.4 8 1
## 20 Lincoln Continental 10.4 8 1
We explain in detail the above steps.
-
windowPartitionBy
creates a window specification objectWindowSpec
that defines the partition. It controls which rows will be in the same partition as the given row. In this case, rows with the same value incyl
will be put in the same partition.orderBy
further defines the ordering - the position a given row is in the partition. The resultingWindowSpec
is returned asws
.
More window specification methods include rangeBetween
, which can define boundaries of the frame by value, and rowsBetween
, which can define the boundaries by row indices.
-
withColumn
appends a Column calledrank
to theSparkDataFrame
.over
returns a windowing column. The first argument is usually a Column returned by window function(s) such asrank()
,lead(carsDF$wt)
. That calculates the corresponding values according to the partitioned-and-ordered table.
User-Defined Function
In SparkR, we support several kinds of user-defined functions (UDFs).
Apply by Partition
dapply
can apply a function to each partition of a SparkDataFrame
. The function to be applied to each partition of the SparkDataFrame
should have only one parameter, a data.frame
corresponding to a partition, and the output should be a data.frame
as well. Schema specifies the row format of the resulting a SparkDataFrame
. It must match to data types of returned value. See here for mapping between R and Spark.
We convert mpg
to kmpg
(kilometers per gallon). carsSubDF
is a SparkDataFrame
with a subset of carsDF
columns.
carsSubDF <- select(carsDF, "model", "mpg")
schema <- "model STRING, mpg DOUBLE, kmpg DOUBLE"
out <- dapply(carsSubDF, function(x) { x <- cbind(x, x$mpg * 1.61) }, schema)
head(collect(out))
## model mpg kmpg
## 1 Mazda RX4 21.0 33.810
## 2 Mazda RX4 Wag 21.0 33.810
## 3 Datsun 710 22.8 36.708
## 4 Hornet 4 Drive 21.4 34.454
## 5 Hornet Sportabout 18.7 30.107
## 6 Valiant 18.1 29.141
Like dapply
, dapplyCollect
can apply a function to each partition of a SparkDataFrame
and collect the result back. The output of the function should be a data.frame
, but no schema is required in this case. Note that dapplyCollect
can fail if the output of the UDF on all partitions cannot be pulled into the driver’s memory.
out <- dapplyCollect(
carsSubDF,
function(x) {
x <- cbind(x, "kmpg" = x$mpg * 1.61)
})
head(out, 3)
## model mpg kmpg
## 1 Mazda RX4 21.0 33.810
## 2 Mazda RX4 Wag 21.0 33.810
## 3 Datsun 710 22.8 36.708
Apply by Group
gapply
can apply a function to each group of a SparkDataFrame
. The function is to be applied to each group of the SparkDataFrame
and should have only two parameters: grouping key and R data.frame
corresponding to that key. The groups are chosen from SparkDataFrames
column(s). The output of function should be a data.frame
. Schema specifies the row format of the resulting SparkDataFrame
. It must represent R function’s output schema on the basis of Spark data types. The column names of the returned data.frame
are set by user. See here for mapping between R and Spark.
schema <- structType(structField("cyl", "double"), structField("max_mpg", "double"))
result <- gapply(
carsDF,
"cyl",
function(key, x) {
y <- data.frame(key, max(x$mpg))
},
schema)
head(arrange(result, "max_mpg", decreasing = TRUE))
## cyl max_mpg
## 1 4 33.9
## 2 6 21.4
## 3 8 19.2
Like gapply
, gapplyCollect
can apply a function to each partition of a SparkDataFrame
and collect the result back to R data.frame
. The output of the function should be a data.frame
but no schema is required in this case. Note that gapplyCollect
can fail if the output of the UDF on all partitions cannot be pulled into the driver’s memory.
result <- gapplyCollect(
carsDF,
"cyl",
function(key, x) {
y <- data.frame(key, max(x$mpg))
colnames(y) <- c("cyl", "max_mpg")
y
})
head(result[order(result$max_mpg, decreasing = TRUE), ])
## cyl max_mpg
## 1 4 33.9
## 2 6 21.4
## 3 8 19.2
Distribute Local Functions
Similar to lapply
in native R, spark.lapply
runs a function over a list of elements and distributes the computations with Spark. spark.lapply
works in a manner that is similar to doParallel
or lapply
to elements of a list. The results of all the computations should fit in a single machine. If that is not the case you can do something like df <- createDataFrame(list)
and then use dapply
.
We use svm
in package e1071
as an example. We use all default settings except for varying costs of constraints violation. spark.lapply
can train those different models in parallel.
costs <- exp(seq(from = log(1), to = log(1000), length.out = 5))
train <- function(cost) {
stopifnot(requireNamespace("e1071", quietly = TRUE))
model <- e1071::svm(Species ~ ., data = iris, cost = cost)
summary(model)
}
Return a list of model’s summaries.
model.summaries <- spark.lapply(costs, train)
class(model.summaries)
## [1] "list"
To avoid lengthy display, we only present the partial result of the second fitted model. You are free to inspect other models as well.
print(model.summaries[[2]])
## $call
## svm(formula = Species ~ ., data = iris, cost = cost)
##
## $type
## [1] 0
##
## $kernel
## [1] 2
##
## $cost
## [1] 5.623413
##
## $degree
## [1] 3
##
## $gamma
## [1] 0.25
##
## $coef0
## [1] 0
##
## $nu
## [1] 0.5
##
## $epsilon
## [1] 0.1
##
## $sparse
## [1] FALSE
##
## $scaled
## [1] TRUE TRUE TRUE TRUE
##
## $x.scale
## $x.scale$`scaled:center`
## Sepal.Length Sepal.Width Petal.Length Petal.Width
## 5.843333 3.057333 3.758000 1.199333
##
## $x.scale$`scaled:scale`
## Sepal.Length Sepal.Width Petal.Length Petal.Width
## 0.8280661 0.4358663 1.7652982 0.7622377
##
##
## $y.scale
## NULL
##
## $nclasses
## [1] 3
##
## $levels
## [1] "setosa" "versicolor" "virginica"
##
## $tot.nSV
## [1] 35
##
## $nSV
## [1] 6 15 14
##
## $labels
## [1] 1 2 3
##
## $SV
## Sepal.Length Sepal.Width Petal.Length Petal.Width
## 14 -1.86378030 -0.13153881 -1.5056946 -1.4422448
## 16 -0.17309407 3.08045544 -1.2791040 -1.0486668
## 21 -0.53538397 0.78617383 -1.1658087 -1.3110521
## 23 -1.50149039 1.24503015 -1.5623422 -1.3110521
## 24 -0.89767388 0.55674567 -1.1658087 -0.9174741
## 42 -1.62225369 -1.73753594 -1.3923993 -1.1798595
## 51 1.39682886 0.32731751 0.5336209 0.2632600
## 53 1.27606556 0.09788935 0.6469162 0.3944526
## 54 -0.41462067 -1.73753594 0.1370873 0.1320673
## 55 0.79301235 -0.59039513 0.4769732 0.3944526
## [ reached getOption("max.print") -- omitted 25 rows ]
##
## $index
## [1] 14 16 21 23 24 42 51 53 54 55 58 61 69 71 73 78 79 84 85
## [20] 86 99 107 111 119 120 124 127 128 130 132 134 135 139 149 150
##
## $rho
## [1] -0.10346530 0.12160294 -0.09540346
##
## $compprob
## [1] FALSE
##
## $probA
## NULL
##
## $probB
## NULL
##
## $sigma
## NULL
##
## $coefs
## [,1] [,2]
## [1,] 0.00000000 0.06561739
## [2,] 0.76813720 0.93378721
## [3,] 0.00000000 0.12123270
## [4,] 0.00000000 0.31170741
## [5,] 1.11614066 0.46397392
## [6,] 1.88141600 1.10392128
## [7,] -0.55872622 0.00000000
## [8,] 0.00000000 5.62341325
## [9,] 0.00000000 0.27711792
## [10,] 0.00000000 5.28440007
## [11,] -1.06596713 0.00000000
## [12,] -0.57076709 1.09019756
## [13,] -0.03365904 5.62341325
## [14,] 0.00000000 5.62341325
## [15,] 0.00000000 5.62341325
## [16,] 0.00000000 5.62341325
## [17,] 0.00000000 4.70398738
## [18,] 0.00000000 5.62341325
## [19,] 0.00000000 4.97981371
## [20,] -0.77497987 0.00000000
## [ reached getOption("max.print") -- omitted 15 rows ]
##
## $na.action
## NULL
##
## $fitted
## 1 2 3 4 5 6 7 8 9 10 11
## setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa
## 12 13 14 15 16 17 18 19 20 21 22
## setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa
## 23 24 25 26 27 28 29 30 31 32 33
## setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa setosa
## 34 35 36 37 38 39 40
## setosa setosa setosa setosa setosa setosa setosa
## [ reached getOption("max.print") -- omitted 110 entries ]
## Levels: setosa versicolor virginica
##
## $decision.values
## setosa/versicolor setosa/virginica versicolor/virginica
## 1 1.1911739 1.0908424 1.1275805
## 2 1.1336557 1.0619543 1.3260964
## 3 1.2085065 1.0698101 1.0511345
## 4 1.1646153 1.0505915 1.0806874
## 5 1.1880814 1.0950348 0.9542815
## 6 1.0990761 1.0984626 0.9326361
## 7 1.1573474 1.0343287 0.9726843
## 8 1.1851598 1.0815750 1.2206802
## 9 1.1673499 1.0406734 0.8837945
## 10 1.1629911 1.0560925 1.2430067
## 11 1.1339282 1.0803946 1.0338357
## 12 1.1724182 1.0641469 1.1190423
## 13 1.1827355 1.0667956 1.1414844
## [ reached getOption("max.print") -- omitted 137 rows ]
##
## $terms
## Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width
## attr(,"variables")
## list(Species, Sepal.Length, Sepal.Width, Petal.Length, Petal.Width)
## attr(,"factors")
## Sepal.Length Sepal.Width Petal.Length Petal.Width
## Species 0 0 0 0
## Sepal.Length 1 0 0 0
## Sepal.Width 0 1 0 0
## Petal.Length 0 0 1 0
## Petal.Width 0 0 0 1
## attr(,"term.labels")
## [1] "Sepal.Length" "Sepal.Width" "Petal.Length" "Petal.Width"
## attr(,"order")
## [1] 1 1 1 1
## attr(,"intercept")
## [1] 0
## attr(,"response")
## [1] 1
## attr(,".Environment")
## <environment: 0x55dea9b744e0>
## attr(,"predvars")
## list(Species, Sepal.Length, Sepal.Width, Petal.Length, Petal.Width)
## attr(,"dataClasses")
## Species Sepal.Length Sepal.Width Petal.Length Petal.Width
## "factor" "numeric" "numeric" "numeric" "numeric"
##
## attr(,"class")
## [1] "summary.svm"
SQL Queries
A SparkDataFrame
can also be registered as a temporary view in Spark SQL so that one can run SQL queries over its data. The sql function enables applications to run SQL queries programmatically and returns the result as a SparkDataFrame
.
people <- read.df(paste0(sparkR.conf("spark.home"),
"/examples/src/main/resources/people.json"), "json")
Register this SparkDataFrame
as a temporary view.
createOrReplaceTempView(people, "people")
SQL statements can be run using the sql method.
## name
## 1 Justin
Machine Learning
SparkR supports the following machine learning models and algorithms.
Classification
Linear Support Vector Machine (SVM) Classifier
Logistic Regression
Multilayer Perceptron (MLP)
Naive Bayes
Factorization Machines (FM) Classifier
Regression
Accelerated Failure Time (AFT) Survival Model
Generalized Linear Model (GLM)
Isotonic Regression
Linear Regression
Factorization Machines (FM) Regressor
Clustering
Bisecting \(k\)-means
Gaussian Mixture Model (GMM)
\(k\)-means Clustering
Latent Dirichlet Allocation (LDA)
Power Iteration Clustering (PIC)
R Formula
For most above, SparkR supports R formula operators, including ~
, .
, :
, +
and -
for model fitting. This makes it a similar experience as using R functions.
Training and Test Sets
We can easily split SparkDataFrame
into random training and test sets by the randomSplit
function. It returns a list of split SparkDataFrames
with provided weights
. We use carsDF
as an example and want to have about \(70%\) training data and \(30%\) test data.
splitDF_list <- randomSplit(carsDF, c(0.7, 0.3), seed = 0)
carsDF_train <- splitDF_list[[1]]
carsDF_test <- splitDF_list[[2]]
count(carsDF_train)
## [1] 24
head(carsDF_train)
## model mpg cyl disp hp drat wt qsec vs am gear carb
## 1 Cadillac Fleetwood 10.4 8 472 205 2.93 5.250 17.98 0 0 3 4
## 2 Camaro Z28 13.3 8 350 245 3.73 3.840 15.41 0 0 3 4
## 3 Chrysler Imperial 14.7 8 440 230 3.23 5.345 17.42 0 0 3 4
## 4 Dodge Challenger 15.5 8 318 150 2.76 3.520 16.87 0 0 3 2
## 5 Duster 360 14.3 8 360 245 3.21 3.570 15.84 0 0 3 4
## 6 Ferrari Dino 19.7 6 145 175 3.62 2.770 15.50 0 1 5 6
count(carsDF_test)
## [1] 8
head(carsDF_test)
## model mpg cyl disp hp drat wt qsec vs am gear carb
## 1 AMC Javelin 15.2 8 304.0 150 3.15 3.435 17.30 0 0 3 2
## 2 Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1
## 3 Fiat 128 32.4 4 78.7 66 4.08 2.200 19.47 1 1 4 1
## 4 Merc 240D 24.4 4 146.7 62 3.69 3.190 20.00 1 0 4 2
## 5 Merc 280 19.2 6 167.6 123 3.92 3.440 18.30 1 0 4 4
## 6 Toyota Corolla 33.9 4 71.1 65 4.22 1.835 19.90 1 1 4 1
Models and Algorithms
Linear Support Vector Machine (SVM) Classifier
Linear Support Vector Machine (SVM) classifier is an SVM classifier with linear kernels. This is a binary classifier. We use a simple example to show how to use spark.svmLinear
for binary classification.
# load training data and create a DataFrame
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
# fit a Linear SVM classifier model
model <- spark.svmLinear(training, Survived ~ ., regParam = 0.01, maxIter = 10)
summary(model)
## $coefficients
## Estimate
## (Intercept) 0.993131388
## Class_1st -0.386500359
## Class_2nd -0.622627816
## Class_3rd -0.204446602
## Sex_Female -0.589950309
## Age_Adult 0.741676902
## Freq -0.006582887
##
## $numClasses
## [1] 2
##
## $numFeatures
## [1] 6
Predict values on training data
prediction <- predict(model, training)
head(select(prediction, "Class", "Sex", "Age", "Freq", "Survived", "prediction"))
## Class Sex Age Freq Survived prediction
## 1 1st Male Child 0 No Yes
## 2 2nd Male Child 0 No Yes
## 3 3rd Male Child 35 No Yes
## 4 Crew Male Child 0 No Yes
## 5 1st Female Child 0 No Yes
## 6 2nd Female Child 0 No No
Logistic Regression
Logistic regression is a widely-used model when the response is categorical. It can be seen as a special case of the Generalized Linear Predictive Model. We provide spark.logit
on top of spark.glm
to support logistic regression with advanced hyper-parameters. It supports both binary and multiclass classification with elastic-net regularization and feature standardization, similar to glmnet
.
We use a simple example to demonstrate spark.logit
usage. In general, there are three steps of using spark.logit
: 1). Create a dataframe from a proper data source; 2). Fit a logistic regression model using spark.logit
with a proper parameter setting; and 3). Obtain the coefficient matrix of the fitted model using summary
and use the model for prediction with predict
.
Binomial logistic regression
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
model <- spark.logit(training, Survived ~ ., regParam = 0.04741301)
summary(model)
## $coefficients
## Estimate
## (Intercept) 0.2255014282
## Class_1st -0.1338856652
## Class_2nd -0.1479826947
## Class_3rd 0.0005674937
## Sex_Female -0.2011183871
## Age_Adult 0.3263186885
## Freq -0.0033111157
Predict values on training data
fitted <- predict(model, training)
head(select(fitted, "Class", "Sex", "Age", "Freq", "Survived", "prediction"))
## Class Sex Age Freq Survived prediction
## 1 1st Male Child 0 No Yes
## 2 2nd Male Child 0 No Yes
## 3 3rd Male Child 35 No Yes
## 4 Crew Male Child 0 No Yes
## 5 1st Female Child 0 No No
## 6 2nd Female Child 0 No No
Multinomial logistic regression against three classes
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
# Note in this case, Spark infers it is multinomial logistic regression, so family = "multinomial" is optional.
model <- spark.logit(training, Class ~ ., regParam = 0.07815179)
summary(model)
## $coefficients
## 1st 2nd 3rd Crew
## (Intercept) 0.051662845 0.062998145 -0.039083689 -0.075577300
## Sex_Female -0.088030587 -0.102528148 0.059233106 0.131325629
## Age_Adult 0.141935316 0.169492058 -0.102562719 -0.208864654
## Survived_No 0.052721020 0.057980057 -0.029408423 -0.081292653
## Freq -0.001555912 -0.001970377 0.001303836 0.002222453
Multilayer Perceptron
Multilayer perceptron classifier (MLPC) is a classifier based on the feedforward artificial neural network. MLPC consists of multiple layers of nodes. Each layer is fully connected to the next layer in the network. Nodes in the input layer represent the input data. All other nodes map inputs to outputs by a linear combination of the inputs with the node’s weights \(w\) and bias \(b\) and applying an activation function. This can be written in matrix form for MLPC with \(K+1\) layers as follows: \[ y(x)=f_K(\ldots f_2(w_2^T f_1(w_1^T x + b_1) + b_2) \ldots + b_K). \]
Nodes in intermediate layers use sigmoid (logistic) function: \[ f(z_i) = \frac{1}{1+e^{-z_i}}. \]
Nodes in the output layer use softmax function: \[ f(z_i) = \frac{e^{z_i}}{\sum_{k=1}^N e^{z_k}}. \]
The number of nodes \(N\) in the output layer corresponds to the number of classes.
MLPC employs backpropagation for learning the model. We use the logistic loss function for optimization and L-BFGS as an optimization routine.
spark.mlp
requires at least two columns in data
: one named "label"
and the other one "features"
. The "features"
column should be in libSVM-format.
We use Titanic data set to show how to use spark.mlp
in classification.
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
# fit a Multilayer Perceptron Classification Model
model <- spark.mlp(training, Survived ~ Age + Sex, blockSize = 128, layers = c(2, 2), solver = "l-bfgs", maxIter = 100, tol = 0.5, stepSize = 1, seed = 1, initialWeights = c( 0, 0, 5, 5, 9, 9))
To avoid lengthy display, we only present partial results of the model summary. You can check the full result from your sparkR shell.
# check the summary of the fitted model
summary(model)
## $numOfInputs
## [1] 2
##
## $numOfOutputs
## [1] 2
##
## $layers
## [1] 2 2
##
## $weights
## $weights[[1]]
## [1] 0
##
## $weights[[2]]
## [1] 0
##
## $weights[[3]]
## [1] 5
##
## $weights[[4]]
## [1] 5
##
## $weights[[5]]
## [1] 9
##
## $weights[[6]]
## [1] 9
# make predictions use the fitted model
predictions <- predict(model, training)
head(select(predictions, predictions$prediction))
## prediction
## 1 No
## 2 No
## 3 No
## 4 No
## 5 No
## 6 No
Naive Bayes
Naive Bayes model assumes independence among the features. spark.naiveBayes
fits a Bernoulli naive Bayes model against a SparkDataFrame. The data should be all categorical. These models are often used for document classification.
titanic <- as.data.frame(Titanic)
titanicDF <- createDataFrame(titanic[titanic$Freq > 0, -5])
naiveBayesModel <- spark.naiveBayes(titanicDF, Survived ~ Class + Sex + Age)
summary(naiveBayesModel)
## $apriori
## Yes No
## [1,] 0.5769231 0.4230769
##
## $tables
## Class_3rd Class_1st Class_2nd Sex_Female Age_Adult
## Yes 0.3125 0.3125 0.3125 0.5 0.5625
## No 0.4166667 0.25 0.25 0.5 0.75
naiveBayesPrediction <- predict(naiveBayesModel, titanicDF)
head(select(naiveBayesPrediction, "Class", "Sex", "Age", "Survived", "prediction"))
## Class Sex Age Survived prediction
## 1 3rd Male Child No Yes
## 2 3rd Female Child No Yes
## 3 1st Male Adult No Yes
## 4 2nd Male Adult No Yes
## 5 3rd Male Adult No No
## 6 Crew Male Adult No Yes
Factorization Machines Classifier
Factorization Machines for classification problems.
For background and details about the implementation of factorization machines, refer to the Factorization Machines section.
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
model <- spark.fmClassifier(training, Survived ~ Age + Sex)
summary(model)
## $coefficients
## Estimate
## (Intercept) 0.0064275991
## Age_Adult 0.0001294448
## Sex_Female 0.0001294448
##
## $factors
## [,1] [,2] [,3] [,4] [,5] [,6]
## [1,] -0.3256224 0.11912568 0.1460235 0.1620567 0.13153516 0.06403695
## [2,] -0.1382155 -0.03658261 0.1717808 -0.1602241 -0.08446129 -0.19287098
## [,7] [,8]
## [1,] -0.03292446 -0.05166818
## [2,] 0.19252571 0.06237194
##
## $numClasses
## [1] 2
##
## $numFeatures
## [1] 2
##
## $factorSize
## [1] 8
## prediction
## 1 Yes
## 2 Yes
## 3 Yes
## 4 Yes
## 5 Yes
## 6 Yes
Accelerated Failure Time Survival Model
Survival analysis studies the expected duration of time until an event happens, and often the relationship with risk factors or treatment taken on the subject. In contrast to standard regression analysis, survival modeling has to deal with special characteristics in the data including non-negative survival time and censoring.
Accelerated Failure Time (AFT) model is a parametric survival model for censored data that assumes the effect of a covariate is to accelerate or decelerate the life course of an event by some constant. For more information, refer to the Wikipedia page AFT Model and the references there. Different from a Proportional Hazards Model designed for the same purpose, the AFT model is easier to parallelize because each instance contributes to the objective function independently.
library(survival)
ovarianDF <- createDataFrame(ovarian)
aftModel <- spark.survreg(ovarianDF, Surv(futime, fustat) ~ ecog_ps + rx)
summary(aftModel)
## $coefficients
## Value
## (Intercept) 6.8966910
## ecog_ps -0.3850414
## rx 0.5286455
## Log(scale) -0.1234429
## futime fustat age resid_ds rx ecog_ps label prediction
## 1 59 1 72.3315 2 1 1 59 1141.724
## 2 115 1 74.4932 2 1 1 115 1141.724
## 3 156 1 66.4658 2 1 2 156 776.855
## 4 421 0 53.3644 2 2 1 421 1937.087
## 5 431 1 50.3397 2 1 1 431 1141.724
## 6 448 0 56.4301 1 1 2 448 776.855
Generalized Linear Model
The main function is spark.glm
. The following families and link functions are supported. The default is gaussian.
Family | Link Function |
---|---|
gaussian | identity, log, inverse |
binomial | logit, probit, cloglog (complementary log-log) |
poisson | log, identity, sqrt |
gamma | inverse, identity, log |
tweedie | power link function |
There are three ways to specify the family
argument.
Family name as a character string, e.g.
family = "gaussian"
.Family function, e.g.
family = binomial
.Result returned by a family function, e.g.
family = poisson(link = log)
.- Note that there are two ways to specify the tweedie family:
- Set
family = "tweedie"
and specify thevar.power
andlink.power
- When package
statmod
is loaded, the tweedie family is specified using the family definition therein, i.e.,tweedie()
.
- Set
For more information regarding the families and their link functions, see the Wikipedia page Generalized Linear Model.
We use the mtcars
dataset as an illustration. The corresponding SparkDataFrame
is carsDF
. After fitting the model, we print out a summary and see the fitted values by making predictions on the original dataset. We can also pass into a new SparkDataFrame
of same schema to predict on new data.
##
## Deviance Residuals:
## (Note: These are approximate quantiles with relative error <= 0.01)
## Min 1Q Median 3Q Max
## -3.9410 -1.6499 -0.3267 1.0373 5.8538
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 37.227270 1.5987875 23.2847 0.0000e+00
## wt -3.877831 0.6327335 -6.1287 1.1196e-06
## hp -0.031773 0.0090297 -3.5187 1.4512e-03
##
## (Dispersion parameter for gaussian family taken to be 6.725785)
##
## Null deviance: 1126.05 on 31 degrees of freedom
## Residual deviance: 195.05 on 29 degrees of freedom
## AIC: 156.7
##
## Number of Fisher Scoring iterations: 1
When doing prediction, a new column called prediction
will be appended. Let’s look at only a subset of columns here.
gaussianFitted <- predict(gaussianGLM, carsDF)
head(select(gaussianFitted, "model", "prediction", "mpg", "wt", "hp"))
## model prediction mpg wt hp
## 1 Mazda RX4 23.57233 21.0 2.620 110
## 2 Mazda RX4 Wag 22.58348 21.0 2.875 110
## 3 Datsun 710 25.27582 22.8 2.320 93
## 4 Hornet 4 Drive 21.26502 21.4 3.215 110
## 5 Hornet Sportabout 18.32727 18.7 3.440 175
## 6 Valiant 20.47382 18.1 3.460 105
The following is the same fit using the tweedie family:
tweedieGLM1 <- spark.glm(carsDF, mpg ~ wt + hp, family = "tweedie", var.power = 0.0)
summary(tweedieGLM1)
##
## Deviance Residuals:
## (Note: These are approximate quantiles with relative error <= 0.01)
## Min 1Q Median 3Q Max
## -3.9410 -1.6499 -0.3267 1.0373 5.8538
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 37.227270 1.5987875 23.2847 0.0000e+00
## wt -3.877831 0.6327335 -6.1287 1.1196e-06
## hp -0.031773 0.0090297 -3.5187 1.4512e-03
##
## (Dispersion parameter for tweedie family taken to be 6.725785)
##
## Null deviance: 1126.05 on 31 degrees of freedom
## Residual deviance: 195.05 on 29 degrees of freedom
## AIC: 156.7
##
## Number of Fisher Scoring iterations: 1
We can try other distributions in the tweedie family, for example, a compound Poisson distribution with a log link:
tweedieGLM2 <- spark.glm(carsDF, mpg ~ wt + hp, family = "tweedie",
var.power = 1.2, link.power = 0.0)
summary(tweedieGLM2)
##
## Deviance Residuals:
## (Note: These are approximate quantiles with relative error <= 0.01)
## Min 1Q Median 3Q Max
## -0.58074 -0.25335 -0.09892 0.18608 0.82717
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 3.8500849 0.06698272 57.4788 0.0000e+00
## wt -0.2018426 0.02897283 -6.9666 1.1691e-07
## hp -0.0016248 0.00041603 -3.9054 5.1697e-04
##
## (Dispersion parameter for tweedie family taken to be 0.1340111)
##
## Null deviance: 29.8820 on 31 degrees of freedom
## Residual deviance: 3.7739 on 29 degrees of freedom
## AIC: NA
##
## Number of Fisher Scoring iterations: 4
Isotonic Regression
spark.isoreg
fits an Isotonic Regression model against a SparkDataFrame
. It solves a weighted univariate a regression problem under a complete order constraint. Specifically, given a set of real observed responses \(y_1, \ldots, y_n\), corresponding real features \(x_1, \ldots, x_n\), and optionally positive weights \(w_1, \ldots, w_n\), we want to find a monotone (piecewise linear) function \(f\) to minimize \[
\ell(f) = \sum_{i=1}^n w_i (y_i - f(x_i))^2.
\]
There are a few more arguments that may be useful.
weightCol
: a character string specifying the weight column.isotonic
: logical value indicating whether the output sequence should be isotonic/increasing (TRUE
) or antitonic/decreasing (FALSE
).featureIndex
: the index of the feature on the right hand side of the formula if it is a vector column (default: 0), no effect otherwise.
We use an artificial example to show the use.
y <- c(3.0, 6.0, 8.0, 5.0, 7.0)
x <- c(1.0, 2.0, 3.5, 3.0, 4.0)
w <- rep(1.0, 5)
data <- data.frame(y = y, x = x, w = w)
df <- createDataFrame(data)
isoregModel <- spark.isoreg(df, y ~ x, weightCol = "w")
isoregFitted <- predict(isoregModel, df)
head(select(isoregFitted, "x", "y", "prediction"))
## x y prediction
## 1 1.0 3 3.0
## 2 2.0 6 5.5
## 3 3.5 8 7.5
## 4 3.0 5 5.5
## 5 4.0 7 7.5
In the prediction stage, based on the fitted monotone piecewise function, the rules are:
If the prediction input exactly matches a training feature then associated prediction is returned. In case there are multiple predictions with the same feature then one of them is returned. Which one is undefined.
If the prediction input is lower or higher than all training features then prediction with lowest or highest feature is returned respectively. In case there are multiple predictions with the same feature then the lowest or highest is returned respectively.
If the prediction input falls between two training features then prediction is treated as piecewise linear function and interpolated value is calculated from the predictions of the two closest features. In case there are multiple values with the same feature then the same rules as in previous point are used.
For example, when the input is \(3.2\), the two closest feature values are \(3.0\) and \(3.5\), then predicted value would be a linear interpolation between the predicted values at \(3.0\) and \(3.5\).
newDF <- createDataFrame(data.frame(x = c(1.5, 3.2)))
head(predict(isoregModel, newDF))
## x prediction
## 1 1.5 4.25
## 2 3.2 6.30
Linear Regression
Linear regression model.
## $coefficients
## Estimate
## (Intercept) 37.22727012
## wt -3.87783074
## hp -0.03177295
##
## $numFeatures
## [1] 2
## prediction
## 1 23.57233
## 2 22.58348
## 3 25.27582
## 4 21.26502
## 5 18.32727
## 6 20.47382
Factorization Machines Regressor
Factorization Machines for regression problems.
For background and details about the implementation of factorization machines, refer to the Factorization Machines section.
model <- spark.fmRegressor(carsDF, mpg ~ wt + hp)
summary(model)
## $coefficients
## Estimate
## (Intercept) 0.1518559
## wt 3.6472555
## hp 2.8026828
##
## $factors
## [,1] [,2] [,3] [,4] [,5] [,6]
## [1,] 0.1424420 -0.1178110 -0.3970272 -0.4696695 0.400288 0.3690930
## [2,] -0.1626185 0.1512138 0.3690435 0.4076975 -0.625752 -0.3715109
## [,7] [,8]
## [1,] 0.03472468 -0.1703219
## [2,] -0.02109148 -0.2006249
##
## $numFeatures
## [1] 2
##
## $factorSize
## [1] 8
## prediction
## 1 106.70996
## 2 87.07526
## 3 111.07931
## 4 60.89565
## 5 61.81374
## 6 40.70095
Decision Tree
spark.decisionTree
fits a decision tree classification or regression model on a SparkDataFrame
. Users can call summary
to get a summary of the fitted model, predict
to make predictions, and write.ml
/read.ml
to save/load fitted models.
We use the Titanic
dataset to train a decision tree and make predictions:
t <- as.data.frame(Titanic)
df <- createDataFrame(t)
dtModel <- spark.decisionTree(df, Survived ~ ., type = "classification", maxDepth = 2)
summary(dtModel)
## Formula: Survived ~ .
## Number of features: 6
## Features: Class_1st Class_2nd Class_3rd Sex_Female Age_Adult Freq
## Feature importances: (6,[5],[1.0])
## Max Depth: 2
## DecisionTreeClassificationModel: uid=dtc_4099e8f5c154, depth=2, numNodes=5, numClasses=2, numFeatures=6
## If (feature 5 <= 4.5)
## Predict: 0.0
## Else (feature 5 > 4.5)
## If (feature 5 <= 84.5)
## Predict: 1.0
## Else (feature 5 > 84.5)
## Predict: 0.0
##
predictions <- predict(dtModel, df)
head(select(predictions, "Class", "Sex", "Age", "Freq", "Survived", "prediction"))
## Class Sex Age Freq Survived prediction
## 1 1st Male Child 0 No No
## 2 2nd Male Child 0 No No
## 3 3rd Male Child 35 No Yes
## 4 Crew Male Child 0 No No
## 5 1st Female Child 0 No No
## 6 2nd Female Child 0 No No
Gradient-Boosted Trees
spark.gbt
fits a gradient-boosted tree classification or regression model on a SparkDataFrame
. Users can call summary
to get a summary of the fitted model, predict
to make predictions, and write.ml
/read.ml
to save/load fitted models.
We use the Titanic
dataset to train a gradient-boosted tree and make predictions:
t <- as.data.frame(Titanic)
df <- createDataFrame(t)
gbtModel <- spark.gbt(df, Survived ~ ., type = "classification", maxDepth = 2, maxIter = 2)
summary(gbtModel)
## Formula: Survived ~ .
## Number of features: 6
## Features: Class_1st Class_2nd Class_3rd Sex_Female Age_Adult Freq
## Feature importances: (6,[1,2,5],[0.03336902858878361,0.16099525743106016,0.8056357139801562])
## Max Depth: 2
## Number of trees: 2
## Tree weights: 1 0.1
## GBTClassificationModel: uid = gbtc_f8afbdc52cba, numTrees=2, numClasses=2, numFeatures=6
## Tree 0 (weight 1.0):
## If (feature 5 <= 4.5)
## If (feature 1 in {1.0})
## Predict: -1.0
## Else (feature 1 not in {1.0})
## Predict: -0.3333333333333333
## Else (feature 5 > 4.5)
## If (feature 5 <= 84.5)
## Predict: 0.5714285714285714
## Else (feature 5 > 84.5)
## Predict: -0.42857142857142855
## Tree 1 (weight 0.1):
## If (feature 2 in {1.0})
## If (feature 5 <= 15.5)
## Predict: 0.9671846896296403
## Else (feature 5 > 15.5)
## Predict: -1.0857923804083338
## Else (feature 2 not in {1.0})
## If (feature 5 <= 13.5)
## Predict: -0.08651035613926407
## Else (feature 5 > 13.5)
## Predict: 0.6566673506774614
##
predictions <- predict(gbtModel, df)
head(select(predictions, "Class", "Sex", "Age", "Freq", "Survived", "prediction"))
## Class Sex Age Freq Survived prediction
## 1 1st Male Child 0 No No
## 2 2nd Male Child 0 No No
## 3 3rd Male Child 35 No Yes
## 4 Crew Male Child 0 No No
## 5 1st Female Child 0 No No
## 6 2nd Female Child 0 No No
Random Forest
spark.randomForest
fits a random forest classification or regression model on a SparkDataFrame
. Users can call summary
to get a summary of the fitted model, predict
to make predictions, and write.ml
/read.ml
to save/load fitted models.
In the following example, we use the Titanic
dataset to train a random forest and make predictions:
t <- as.data.frame(Titanic)
df <- createDataFrame(t)
rfModel <- spark.randomForest(df, Survived ~ ., type = "classification", maxDepth = 2, numTrees = 2)
summary(rfModel)
## Formula: Survived ~ .
## Number of features: 6
## Features: Class_1st Class_2nd Class_3rd Sex_Female Age_Adult Freq
## Feature importances: (6,[3,4,5],[0.17058779274099098,0.09676977311565654,0.7326424341433525])
## Max Depth: 2
## Number of trees: 2
## Tree weights: 1 1
## RandomForestClassificationModel: uid=rfc_752b6d4a7c73, numTrees=2, numClasses=2, numFeatures=6
## Tree 0 (weight 1.0):
## If (feature 4 in {0.0})
## If (feature 3 in {0.0})
## Predict: 0.0
## Else (feature 3 not in {0.0})
## Predict: 1.0
## Else (feature 4 not in {0.0})
## If (feature 5 <= 13.5)
## Predict: 0.0
## Else (feature 5 > 13.5)
## Predict: 1.0
## Tree 1 (weight 1.0):
## If (feature 5 <= 84.5)
## If (feature 5 <= 4.5)
## Predict: 0.0
## Else (feature 5 > 4.5)
## Predict: 1.0
## Else (feature 5 > 84.5)
## Predict: 0.0
##
predictions <- predict(rfModel, df)
head(select(predictions, "Class", "Sex", "Age", "Freq", "Survived", "prediction"))
## Class Sex Age Freq Survived prediction
## 1 1st Male Child 0 No No
## 2 2nd Male Child 0 No No
## 3 3rd Male Child 35 No Yes
## 4 Crew Male Child 0 No No
## 5 1st Female Child 0 No No
## 6 2nd Female Child 0 No No
Bisecting k-Means
spark.bisectingKmeans
is a kind of hierarchical clustering using a divisive (or “top-down”) approach: all observations start in one cluster, and splits are performed recursively as one moves down the hierarchy.
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
model <- spark.bisectingKmeans(training, Class ~ Survived, k = 4)
summary(model)
## $k
## [1] 4
##
## $coefficients
## Survived_No
## 1 0
## 2 1
## 3 0
## 4 1
##
## $size
## $size[[1]]
## [1] 16
##
## $size[[2]]
## [1] 16
##
## $size[[3]]
## [1] 0
##
## $size[[4]]
## [1] 0
##
##
## $cluster
## SparkDataFrame[prediction:int]
##
## $is.loaded
## [1] FALSE
## Class prediction
## 1 1st 1
## 2 2nd 1
## 3 3rd 1
## 4 Crew 1
## 5 1st 1
## 6 2nd 1
Gaussian Mixture Model
spark.gaussianMixture
fits multivariate Gaussian Mixture Model (GMM) against a SparkDataFrame
. Expectation-Maximization (EM) is used to approximate the maximum likelihood estimator (MLE) of the model.
We use a simulated example to demonstrate the usage.
X1 <- data.frame(V1 = rnorm(4), V2 = rnorm(4))
X2 <- data.frame(V1 = rnorm(6, 3), V2 = rnorm(6, 4))
data <- rbind(X1, X2)
df <- createDataFrame(data)
gmmModel <- spark.gaussianMixture(df, ~ V1 + V2, k = 2)
summary(gmmModel)
## $lambda
## [1] 0.4000018 0.5999982
##
## $mu
## $mu[[1]]
## [1] 0.6779989 -0.1129207
##
## $mu[[2]]
## [1] 3.163231 4.551850
##
##
## $sigma
## $sigma[[1]]
## [,1] [,2]
## [1,] 0.8067735 0.8476199
## [2,] 0.8476199 0.9524243
##
## $sigma[[2]]
## [,1] [,2]
## [1,] 0.03492959 0.001296386
## [2,] 0.001296386 0.4321705
##
##
## $loglik
## [1] -16.53407
##
## $posterior
## SparkDataFrame[posterior:array<double>]
##
## $is.loaded
## [1] FALSE
## V1 V2 prediction
## 1 0.5953528 0.1994821 0
## 2 1.2085170 0.1759681 0
## 3 1.6453459 0.8993440 0
## 4 -0.7372627 -1.7265389 0
## 5 3.0104312 4.9300509 1
## 6 2.9178137 4.3606677 1
k-Means Clustering
spark.kmeans
fits a \(k\)-means clustering model against a SparkDataFrame
. As an unsupervised learning method, we don’t need a response variable. Hence, the left hand side of the R formula should be left blank. The clustering is based only on the variables on the right hand side.
kmeansModel <- spark.kmeans(carsDF, ~ mpg + hp + wt, k = 3)
summary(kmeansModel)
## $k
## [1] 3
##
## $coefficients
## mpg hp wt
## 1 24.22353 93.52941 2.599588
## 2 15.80000 178.50000 3.926400
## 3 14.62000 263.80000 3.899000
##
## $size
## $size[[1]]
## [1] 17
##
## $size[[2]]
## [1] 10
##
## $size[[3]]
## [1] 5
##
##
## $cluster
## SparkDataFrame[prediction:int]
##
## $is.loaded
## [1] FALSE
##
## $clusterSize
## [1] 3
kmeansPredictions <- predict(kmeansModel, carsDF)
head(select(kmeansPredictions, "model", "mpg", "hp", "wt", "prediction"), n = 20L)
## model mpg hp wt prediction
## 1 Mazda RX4 21.0 110 2.620 0
## 2 Mazda RX4 Wag 21.0 110 2.875 0
## 3 Datsun 710 22.8 93 2.320 0
## 4 Hornet 4 Drive 21.4 110 3.215 0
## 5 Hornet Sportabout 18.7 175 3.440 1
## 6 Valiant 18.1 105 3.460 0
## 7 Duster 360 14.3 245 3.570 2
## 8 Merc 240D 24.4 62 3.190 0
## 9 Merc 230 22.8 95 3.150 0
## 10 Merc 280 19.2 123 3.440 0
## 11 Merc 280C 17.8 123 3.440 0
## 12 Merc 450SE 16.4 180 4.070 1
## 13 Merc 450SL 17.3 180 3.730 1
## 14 Merc 450SLC 15.2 180 3.780 1
## 15 Cadillac Fleetwood 10.4 205 5.250 1
## 16 Lincoln Continental 10.4 215 5.424 1
## 17 Chrysler Imperial 14.7 230 5.345 2
## 18 Fiat 128 32.4 66 2.200 0
## 19 Honda Civic 30.4 52 1.615 0
## 20 Toyota Corolla 33.9 65 1.835 0
Latent Dirichlet Allocation
spark.lda
fits a Latent Dirichlet Allocation model on a SparkDataFrame
. It is often used in topic modeling in which topics are inferred from a collection of text documents. LDA can be thought of as a clustering algorithm as follows:
Topics correspond to cluster centers, and documents correspond to examples (rows) in a dataset.
Topics and documents both exist in a feature space, where feature vectors are vectors of word counts (bag of words).
Rather than clustering using a traditional distance, LDA uses a function based on a statistical model of how text documents are generated.
To use LDA, we need to specify a features
column in data
where each entry represents a document. There are two options for the column:
character string: This can be a string of the whole document. It will be parsed automatically. Additional stop words can be added in
customizedStopWords
.libSVM: Each entry is a collection of words and will be processed directly.
Two more functions are provided for the fitted model.
spark.posterior
returns aSparkDataFrame
containing a column of posterior probabilities vectors named “topicDistribution”.spark.perplexity
returns the log perplexity of givenSparkDataFrame
, or the log perplexity of the training data if missing argumentdata
.
For more information, see the help document ?spark.lda
.
Let’s look an artificial example.
corpus <- data.frame(features = c(
"1 2 6 0 2 3 1 1 0 0 3",
"1 3 0 1 3 0 0 2 0 0 1",
"1 4 1 0 0 4 9 0 1 2 0",
"2 1 0 3 0 0 5 0 2 3 9",
"3 1 1 9 3 0 2 0 0 1 3",
"4 2 0 3 4 5 1 1 1 4 0",
"2 1 0 3 0 0 5 0 2 2 9",
"1 1 1 9 2 1 2 0 0 1 3",
"4 4 0 3 4 2 1 3 0 0 0",
"2 8 2 0 3 0 2 0 2 7 2",
"1 1 1 9 0 2 2 0 0 3 3",
"4 1 0 0 4 5 1 3 0 1 0"))
corpusDF <- createDataFrame(corpus)
model <- spark.lda(data = corpusDF, k = 5, optimizer = "em")
summary(model)
## $docConcentration
## [1] 11 11 11 11 11
##
## $topicConcentration
## [1] 1.1
##
## $logLikelihood
## [1] -353.2948
##
## $logPerplexity
## [1] 2.676476
##
## $isDistributed
## [1] TRUE
##
## $vocabSize
## [1] 10
##
## $topics
## SparkDataFrame[topic:int, term:array<string>, termWeights:array<double>]
##
## $vocabulary
## [1] "0" "1" "2" "3" "4" "9" "5" "8" "7" "6"
##
## $trainingLogLikelihood
## [1] -239.5629
##
## $logPrior
## [1] -980.2974
posterior <- spark.posterior(model, corpusDF)
head(posterior)
## features topicDistribution
## 1 1 2 6 0 2 3 1 1 0 0 3 0.1972166, 0.1986653, 0.2022007, 0.2006584, 0.2012590
## 2 1 3 0 1 3 0 0 2 0 0 1 0.1989970, 0.1988748, 0.2015960, 0.2006387, 0.1998935
## 3 1 4 1 0 0 4 9 0 1 2 0 0.2020592, 0.2026098, 0.1968834, 0.1987297, 0.1997178
## 4 2 1 0 3 0 0 5 0 2 3 9 0.2004063, 0.1981944, 0.2013008, 0.2006321, 0.1994664
## 5 3 1 1 9 3 0 2 0 0 1 3 0.1971472, 0.1983963, 0.2023570, 0.2011593, 0.2009403
## 6 4 2 0 3 4 5 1 1 1 4 0 0.2020220, 0.2041838, 0.1955390, 0.1997235, 0.1985317
perplexity <- spark.perplexity(model, corpusDF)
perplexity
## [1] 2.676476
Alternating Least Squares
spark.als
learns latent factors in collaborative filtering via alternating least squares.
There are multiple options that can be configured in spark.als
, including rank
, reg
, and nonnegative
. For a complete list, refer to the help file.
ratings <- list(list(0, 0, 4.0), list(0, 1, 2.0), list(1, 1, 3.0), list(1, 2, 4.0),
list(2, 1, 1.0), list(2, 2, 5.0))
df <- createDataFrame(ratings, c("user", "item", "rating"))
model <- spark.als(df, "rating", "user", "item", rank = 10, reg = 0.1, nonnegative = TRUE)
Extract latent factors.
stats <- summary(model)
userFactors <- stats$userFactors
itemFactors <- stats$itemFactors
head(userFactors)
head(itemFactors)
Make predictions.
Power Iteration Clustering
Power Iteration Clustering (PIC) is a scalable graph clustering algorithm. spark.assignClusters
method runs the PIC algorithm and returns a cluster assignment for each input vertex.
df <- createDataFrame(list(list(0L, 1L, 1.0), list(0L, 2L, 1.0),
list(1L, 2L, 1.0), list(3L, 4L, 1.0),
list(4L, 0L, 0.1)),
schema = c("src", "dst", "weight"))
head(spark.assignClusters(df, initMode = "degree", weightCol = "weight"))
## id cluster
## 1 4 1
## 2 0 0
## 3 1 0
## 4 3 1
## 5 2 0
FP-growth
spark.fpGrowth
executes FP-growth algorithm to mine frequent itemsets on a SparkDataFrame
. itemsCol
should be an array of values.
df <- selectExpr(createDataFrame(data.frame(rawItems = c(
"T,R,U", "T,S", "V,R", "R,U,T,V", "R,S", "V,S,U", "U,R", "S,T", "V,R", "V,U,S",
"T,V,U", "R,V", "T,S", "T,S", "S,T", "S,U", "T,R", "V,R", "S,V", "T,S,U"
))), "split(rawItems, ',') AS items")
fpm <- spark.fpGrowth(df, minSupport = 0.2, minConfidence = 0.5)
spark.freqItemsets
method can be used to retrieve a SparkDataFrame
with the frequent itemsets.
head(spark.freqItemsets(fpm))
## items freq
## 1 R 9
## 2 U 8
## 3 U, T 4
## 4 U, V 4
## 5 U, S 4
## 6 T 10
spark.associationRules
returns a SparkDataFrame
with the association rules.
head(spark.associationRules(fpm))
## antecedent consequent confidence lift support
## 1 V R 0.5555556 1.234568 0.25
## 2 S T 0.5454545 1.090909 0.30
## 3 T S 0.6000000 1.090909 0.30
## 4 R V 0.5555556 1.234568 0.25
## 5 U T 0.5000000 1.000000 0.20
## 6 U V 0.5000000 1.111111 0.20
We can make predictions based on the antecedent
.
## items prediction
## 1 T, R, U S, V
## 2 T, S NULL
## 3 V, R NULL
## 4 R, U, T, V S
## 5 R, S T, V
## 6 V, S, U R, T
PrefixSpan
spark.findFrequentSequentialPatterns
method can be used to find the complete set of frequent sequential patterns in the input sequences of itemsets.
df <- createDataFrame(list(list(list(list(1L, 2L), list(3L))),
list(list(list(1L), list(3L, 2L), list(1L, 2L))),
list(list(list(1L, 2L), list(5L))),
list(list(list(6L)))),
schema = c("sequence"))
head(spark.findFrequentSequentialPatterns(df, minSupport = 0.5, maxPatternLength = 5L))
## sequence freq
## 1 1 3
## 2 3 2
## 3 2 3
## 4 1, 2 3
## 5 1, 3 2
Kolmogorov-Smirnov Test
spark.kstest
runs a two-sided, one-sample Kolmogorov-Smirnov (KS) test. Given a SparkDataFrame
, the test compares continuous data in a given column testCol
with the theoretical distribution specified by parameter nullHypothesis
. Users can call summary
to get a summary of the test results.
In the following example, we test whether the Titanic
dataset’s Freq
column follows a normal distribution. We set the parameters of the normal distribution using the mean and standard deviation of the sample.
t <- as.data.frame(Titanic)
df <- createDataFrame(t)
freqStats <- head(select(df, mean(df$Freq), sd(df$Freq)))
freqMean <- freqStats[1]
freqStd <- freqStats[2]
test <- spark.kstest(df, "Freq", "norm", c(freqMean, freqStd))
testSummary <- summary(test)
testSummary
## Kolmogorov-Smirnov test summary:
## degrees of freedom = 0
## statistic = 0.3065126710255011
## pValue = 0.0036336792155329256
## Very strong presumption against null hypothesis: Sample follows theoretical distribution.
Model Persistence
The following example shows how to save/load an ML model in SparkR.
t <- as.data.frame(Titanic)
training <- createDataFrame(t)
gaussianGLM <- spark.glm(training, Freq ~ Sex + Age, family = "gaussian")
# Save and then load a fitted MLlib model
modelPath <- tempfile(pattern = "ml", fileext = ".tmp")
write.ml(gaussianGLM, modelPath)
gaussianGLM2 <- read.ml(modelPath)
# Check model summary
summary(gaussianGLM2)
##
## Saved-loaded model does not support output 'Deviance Residuals'.
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 46.219 35.994 1.2841 0.2092846
## Sex_Female -78.812 41.562 -1.8962 0.0679311
## Age_Adult 123.938 41.562 2.9820 0.0057522
##
## (Dispersion parameter for gaussian family taken to be 13819.52)
##
## Null deviance: 573341 on 31 degrees of freedom
## Residual deviance: 400766 on 29 degrees of freedom
## AIC: 400.7
##
## Number of Fisher Scoring iterations: 1
# Check model prediction
gaussianPredictions <- predict(gaussianGLM2, training)
head(gaussianPredictions)
## Class Sex Age Survived Freq label prediction
## 1 1st Male Child No 0 0 46.21875
## 2 2nd Male Child No 0 0 46.21875
## 3 3rd Male Child No 35 35 46.21875
## 4 Crew Male Child No 0 0 46.21875
## 5 1st Female Child No 0 0 -32.59375
## 6 2nd Female Child No 0 0 -32.59375
unlink(modelPath)
Structured Streaming
SparkR supports the Structured Streaming API.
You can check the Structured Streaming Programming Guide for an introduction to its programming model and basic concepts.
Simple Source and Sink
Spark has a few built-in input sources. As an example, to test with a socket source reading text into words and displaying the computed word counts:
# Create DataFrame representing the stream of input lines from connection
lines <- read.stream("socket", host = hostname, port = port)
# Split the lines into words
words <- selectExpr(lines, "explode(split(value, ' ')) as word")
# Generate running word count
wordCounts <- count(groupBy(words, "word"))
# Start running the query that prints the running counts to the console
query <- write.stream(wordCounts, "console", outputMode = "complete")
Kafka Source
It is simple to read data from Kafka. For more information, see Input Sources supported by Structured Streaming.
topic <- read.stream("kafka",
kafka.bootstrap.servers = "host1:port1,host2:port2",
subscribe = "topic1")
keyvalue <- selectExpr(topic, "CAST(key AS STRING)", "CAST(value AS STRING)")
Operations and Sinks
Most of the common operations on SparkDataFrame
are supported for streaming, including selection, projection, and aggregation. Once you have defined the final result, to start the streaming computation, you will call the write.stream
method setting a sink and outputMode
.
A streaming SparkDataFrame
can be written for debugging to the console, to a temporary in-memory table, or for further processing in a fault-tolerant manner to a File Sink in different formats.
noAggDF <- select(where(deviceDataStreamingDf, "signal > 10"), "device")
# Print new data to console
write.stream(noAggDF, "console")
# Write new data to Parquet files
write.stream(noAggDF,
"parquet",
path = "path/to/destination/dir",
checkpointLocation = "path/to/checkpoint/dir")
# Aggregate
aggDF <- count(groupBy(noAggDF, "device"))
# Print updated aggregations to console
write.stream(aggDF, "console", outputMode = "complete")
# Have all the aggregates in an in memory table. The query name will be the table name
write.stream(aggDF, "memory", queryName = "aggregates", outputMode = "complete")
head(sql("select * from aggregates"))
Advanced Topics
SparkR Object Classes
There are three main object classes in SparkR you may be working with.
-
SparkDataFrame
: the central component of SparkR. It is an S4 class representing distributed collection of data organized into named columns, which is conceptually equivalent to a table in a relational database or a data frame in R. It has two slotssdf
andenv
.-
sdf
stores a reference to the corresponding Spark Dataset in the Spark JVM backend. -
env
saves the meta-information of the object such asisCached
.
It can be created by data import methods or by transforming an existing
SparkDataFrame
. We can manipulateSparkDataFrame
by numerous data processing functions and feed that into machine learning algorithms. -
-
Column
: an S4 class representing a column ofSparkDataFrame
. The slotjc
saves a reference to the correspondingColumn
object in the Spark JVM backend.It can be obtained from a
SparkDataFrame
by$
operator, e.g.,df$col
. More often, it is used together with other functions, for example, withselect
to select particular columns, withfilter
and constructed conditions to select rows, with aggregation functions to compute aggregate statistics for each group. -
GroupedData
: an S4 class representing grouped data created bygroupBy
or by transforming otherGroupedData
. Itssgd
slot saves a reference to aRelationalGroupedDataset
object in the backend.This is often an intermediate object with group information and followed up by aggregation operations.
Architecture
A complete description of architecture can be seen in the references, in particular the paper SparkR: Scaling R Programs with Spark.
Under the hood of SparkR is Spark SQL engine. This avoids the overheads of running interpreted R code, and the optimized SQL execution engine in Spark uses structural information about data and computation flow to perform a bunch of optimizations to speed up the computation.
The main method calls of actual computation happen in the Spark JVM of the driver. We have a socket-based SparkR API that allows us to invoke functions on the JVM from R. We use a SparkR JVM backend that listens on a Netty-based socket server.
Two kinds of RPCs are supported in the SparkR JVM backend: method invocation and creating new objects. Method invocation can be done in two ways.
sparkR.callJMethod
takes a reference to an existing Java object and a list of arguments to be passed on to the method.sparkR.callJStatic
takes a class name for static method and a list of arguments to be passed on to the method.
The arguments are serialized using our custom wire format which is then deserialized on the JVM side. We then use Java reflection to invoke the appropriate method.
To create objects, sparkR.newJObject
is used and then similarly the appropriate constructor is invoked with provided arguments.
Finally, we use a new R class jobj
that refers to a Java object existing in the backend. These references are tracked on the Java side and are automatically garbage collected when they go out of scope on the R side.
References
SparkR: Scaling R Programs with Spark, Shivaram Venkataraman, Zongheng Yang, Davies Liu, Eric Liang, Hossein Falaki, Xiangrui Meng, Reynold Xin, Ali Ghodsi, Michael Franklin, Ion Stoica, and Matei Zaharia. SIGMOD 2016. June 2016.