FPgrowth computing association in pyspark vs scala

问题:

Using :

http://spark.apache.org/docs/1.6.1/mllib-frequent-pattern-mining.html

Python Code:

from pyspark.mllib.fpm import FPGrowth
model = FPGrowth.train(dataframe,0.01,10)

Scala:

import org.apache.spark.mllib.fpm.FPGrowth
import org.apache.spark.rdd.RDD

val data = sc.textFile("data/mllib/sample_fpgrowth.txt")

val transactions: RDD[Array[String]] = data.map(s => s.trim.split(' '))

val fpg = new FPGrowth()
  .setMinSupport(0.2)
  .setNumPartitions(10)
val model = fpg.run(transactions)

model.freqItemsets.collect().foreach { itemset =>
  println(itemset.items.mkString("[", ",", "]") + ", " + itemset.freq)
}

val minConfidence = 0.8
model.generateAssociationRules(minConfidence).collect().foreach { rule =>
  println(
    rule.antecedent.mkString("[", ",", "]")
      + " => " + rule.consequent .mkString("[", ",", "]")
      + ", " + rule.confidence)
}

From code here it shows that scala part doesn't have minimum confidence.

def trainFPGrowthModel(
      data: JavaRDD[java.lang.Iterable[Any]],
      minSupport: Double,
      numPartitions: Int): FPGrowthModel[Any] = {
    val fpg = new FPGrowth()
      .setMinSupport(minSupport)
      .setNumPartitions(numPartitions)

    val model = fpg.run(data.rdd.map(_.asScala.toArray))
    new FPGrowthModelWrapper(model)
  }

How to add minConfidence to generate association rule in case of pyspark? We can see that scala has the example but python does not have the example.

回答1:

Spark >= 2.2

There is a DataFrame base ml API which provides AssociationRules:

from pyspark.ml.fpm import FPGrowth

data = ...

fpm = FPGrowth(minSupport=0.3, minConfidence=0.9).fit(data)
associationRules = fpm.associationRules.

Spark < 2.2

As for now PySpark doesn't support extracting association rules (DataFrame based FPGrowth API with Python support is a work in progress SPARK-1450) but we can easily address that.

First you'll have to install SBT (just go the downloads page) and follow the instructions for your operating system.

Next you'll have to create a simple Scala project with only two files:

.
├── AssociationRulesExtractor.scala
└── build.sbt

You can adjust it later to follow the established directory structure.

Next add following to the build.sbt (adjust Scala version and Spark version to match the one you use):

name := "fpm"

version := "1.0"

scalaVersion := "2.10.6"

val sparkVersion = "1.6.2"

libraryDependencies ++= Seq(
  "org.apache.spark" %% "spark-core" % sparkVersion,
  "org.apache.spark" %% "spark-mllib" % sparkVersion
)

and following to the AssociationRulesExtractor.scala:

package com.example.fpm

import org.apache.spark.mllib.fpm.AssociationRules.Rule
import org.apache.spark.rdd.RDD

object AssociationRulesExtractor {
  def apply(rdd: RDD[Rule[String]]) = {
    rdd.map(rule => Array(
      rule.confidence, rule.javaAntecedent, rule.javaConsequent
    ))
  }
}

Open terminal emulator of your choice, go to the root directory of the project and call:

sbt package

It will generate a jar file in the target directory. For example in Scala 2.10 it will be:

target/scala-2.10/fpm_2.10-1.0.jar

Start PySpark shell or use spark-submit and pass path to the generated jar file as to --driver-class-path:

bin/pyspark --driver-class-path /path/to/fpm_2.10-1.0.jar

In non-local mode:

bin/pyspark --driver-class-path /path/to/fpm_2.10-1.0.jar --jars /path/to/fpm_2.10-1.0.jar

In cluster mode jar should be present on all nodes.

Add some convenience wrappers:

from pyspark import SparkContext
from pyspark.mllib.fpm import FPGrowthModel
from pyspark.mllib.common import _java2py
from collections import namedtuple


rule = namedtuple("Rule", ["confidence", "antecedent", "consequent"])

def generateAssociationRules(model, minConfidence):
    # Get active context
    sc = SparkContext.getOrCreate()

    # Retrieve extractor object
    extractor = sc._gateway.jvm.com.example.fpm.AssociationRulesExtractor

    # Compute rules
    java_rules = model._java_model.generateAssociationRules(minConfidence)

    # Convert rules to Python RDD
    return _java2py(sc, extractor.apply(java_rules)).map(lambda x:rule(*x))

Finally you can use these helpers as a function:

generateAssociationRules(model, 0.9)

or as a method:

FPGrowthModel.generateAssociationRules = generateAssociationRules
model.generateAssociationRules(0.9)

This solution depends on internal PySpark methods so it is not guaranteed that it will be portable between versions.

回答2:

You can generate and get association rules in PySpark using Spark <2.2 with a little bit of py4j code:

# model was produced by FPGrowth.train() method
rules = sorted(model._java_model.generateAssociationRules(0.9).collect(), 
    key=lambda x: x.confidence(), reverse=True)
for rule in rules[:200]:
    # rule variable has confidence(), consequent() and antecedent() 
    # methods for individual value access.
    print rule