基于Spark GraphX 的图形数据分析

为什么需要图计算

• 许多大数据以大规模图或网络的形式呈现
• 许多非图结构的大数据，常会被转换为图模型进行分析
• 图数据结构很好地表达了数据之间的关联性

一.图（Graph）的基本概念

图是由顶点集合(vertex)及顶点间的关系集合（边edge）组成的一种网状数据结构

• 通常表示为二元组：Gragh=（V，E）
• 可以对事物之间的关系建模

应用场景

• 在地图应用中寻找最短路径
• 社交网络关系
• 网页间超链接关系

图的术语

• 顶点（Vertex）
• 边（Edge）

Graph=(V,E)
集合V={
    
    v1,v2,v3}
集合E={
    
    (v1,v2),(v1,v3),(v2,v3)}

• 有向图

G=(V,E)
V={
    
    A,B,C,D,E}
E={
    
    <A,B>,<B,C>,<B,D>,<C,E>,<D,A>,<E,D>}

• 无向图

G=(V,E)
V={
    
    A,B,C,D,E}
E={
    
    (A,B),(A,D),(B,C),(B,D),(C,E),(D,E)}

• 有环图:包含一系列顶点连接的回路（环路）
  

• 无环图:DAG即为有向无环图
  

度：一个顶点所有边的数量

• 出度：指从当前顶点指向其他顶点的边的数量

• 入度：其他顶点指向当前顶点的边的数量
  

• 图的经典表示法:邻接矩阵

1、对于每条边，矩阵中相应单元格值为1
2、对于每个循环，矩阵中相应单元格值为2，方便在行或列上求得顶点度数

二.Spark GraphX 简介

GraphX是Spark提供分布式图计算API
GraphX特点

• 基于内存实现了数据的复用与快速读取
• 通过弹性分布式属性图（Property Graph）统一了图视图与表视图
• 与Spark Streaming、Spark SQL和Spark MLlib等无缝衔接

GraphX核心抽象

• 弹性分布式属性图（Resilient Distributed Property Graph）

• 顶点和边都带属性的有向多重图
  
  

• 一份物理存储，两种视图
  

对Graph视图的所有操作，最终都会转换成其关联的Table视图的RDD操作来完成

三.GraphX API

• Graph[VD,ED]
• VertexRDD[VD]
• EdgeRDD[ED]
• EdgeTriplet[VD,ED]
• Edge：样例类
• VertexId：Long的别名

class Graph[VD, ED] {
    
    
  val vertices: VertexRDD[VD]
  val edges: EdgeRDD[ED]
  val triplets: RDD[EdgeTriplet[VD, ED]]
}

import org.apache.spark.graphx._
val vertices:RDD[(VertexId,Int)]=sc.makeRDD(Seq((1L,1),(2L,2),(3L,3)))
val edges=sc.makeRDD(Seq(Edge(1L,2L,1),Edge(2L,3L,2)))
val graph=Graph(vertices,edges)  //Graph[Int,Int] ?

import org.apache.spark.graphx.GraphLoader
//加载边列表文件创建图，文件每行描述一条边，格式：srcId dstId。顶点与边的属性均为1
val graph = GraphLoader.edgeListFile(sc, 
                                        "file:///opt/spark/data/graphx/followers.txt")

1.属性图应用示例-1

构建用户合作关系属性图

• 顶点属性:用户名.职业
• 边属性:合作关系

val userGraph: Graph[(String, String), String]

2.属性图应用示例-2

构建用户社交网络关系

• 顶点：用户名、年龄
• 边：打call次数

找出大于30岁的用户

• 假设打call超过5次，表示真爱。请找出他（她）们
  

import org.apache.spark.graphx._
val v1 = sc.makeRDD(Array((1L,("Alice",28)),(2L,("Bob",27)),(3L,("Charlie",65)),(4L,("David",42)),(5L,("Ed",55)),(6L,("Fran",50))))
val e1 = sc.makeRDD(Array(Edge(4L,1L,1L),Edge(2L,1L,7L),Edge(2L,4L,2L),Edge(5L,2L,2L),Edge(5L,3L,8L),Edge(5L,6L,3L),Edge(3L,6L,3L),Edge(3L,2L,4L)))
val graph1 =Graph(v1,e1)
//大于30岁的人
graph1.vertices.filter{
    
    case(id,(name,age))=>age>30}.collect
//打call超过5次
graph1.edges.filter{
    
    case(Edge(from,to,call))=>call>5}.collect

3.查看图信息

• 顶点数量
• 边数量
• 度、入度、出度

class Graph[VD, ED] {
    
    
  val numEdges: Long
  val numVertices: Long
  val inDegrees: VertexRDD[Int]
  val outDegrees: VertexRDD[Int]
  val degrees: VertexRDD[Int]
}

4.图的算子

属性算子:类似于RDD的map操作

class Graph[VD, ED] {
    
    
  def mapVertices[VD2](map: (VertexId, VD) => VD2): Graph[VD2, ED]
  def mapEdges[ED2](map: Edge[ED] => ED2): Graph[VD, ED2]
  def mapTriplets[ED2](map: EdgeTriplet[VD, ED] => ED2): Graph[VD, ED2]
}

val t1_graph = tweeter_graph.mapVertices {
    
     case(vertextId, (name, age)) => (vertextId, name) }
val t2_graph = tweeter_graph.mapVertices {
    
     (vertextId, attr) => (vertextId, attr._1) }
val t3_graph = tweeter_graph.mapEdges(e => Edge(e.srcId, e.dstId, e.attr*7.0))

结构算子

class Graph[VD, ED] {
    
    
  def reverse: Graph[VD, ED] //改变边的方向
  //生成满足顶点与边的条件的子图
  def subgraph(epred: EdgeTriplet[VD,ED] => Boolean, 
               vpred: (VertexId, VD) => Boolean): Graph[VD, ED]
  }

val t1_graph = tweeter_graph.reverse
val t2_graph = tweeter_graph.subgraph(vpred=(id,attr)=>attr._2<65)  //attr:(name,age)

Join算子：从外部的RDDs加载数据，修改顶点属性

class Graph[VD, ED] {
    
    
  def joinVertices[U](table: RDD[(VertexId, U)])(map: (VertexId, VD, U) => VD): Graph[VD, ED]
  //RDD中的顶点不匹配时，值为None
  def outerJoinVertices[U, VD2](table: RDD[(VertexId, U)])(map: (VertexId, VD, Option[U]) => VD2)
    : Graph[VD2, ED]
}

val tweeters_comps:RDD[(VertexId,String)]= sc.parallelize(Array((1L, "kgc.cn"), (2L, "berkeley.edu"), (3L, "apache.org")))
val t_graph = tweeter_graph.joinVertices(tweeters_comps)((id, v, cmpy) => (v._1 + " @ " + cmpy, v._2))
t_graph.vertices.collect

val s_graph = tweeter_graph.outerJoinVertices(tweeters_comps)((id, v, cmpy) => (v._1 + " @ " + cmpy, v._2))
s_graph.vertices.collect

5.GraphX API 应用

import org.apache.spark.graphx._
val v1 = sc.makeRDD(Array((1L,("Alice",28)),(2L,("Bob",27)),(3L,("Charlie",65)),(4L,("David",42)),(5L,("Ed",55)),(6L,("Fran",50))))
val e1 = sc.makeRDD(Array(Edge(4L,1L,1),Edge(2L,1L,7),Edge(2L,4L,2),Edge(5L,2L,2),Edge(5L,3L,8),Edge(5L,6L,3),Edge(3L,6L,3),Edge(3L,2L,4)))
val graph1 =Graph(v1,e1)
case class User(name: String, age: Int, inDeg: Int, outDeg: Int)
//修改顶点属性
val initialUserGraph: Graph[User, Int] = tweeter_graph.mapVertices{
    
     
     case (id, (name, age)) => User(name, age, 0, 0) 
}
//将顶点入度、出度存入顶点属性中 
val userGraph = initialUserGraph.outerJoinVertices(initialUserGraph.inDegrees) {
    
    
     case (id, u, inDegOpt) => User(u.name, u.age, inDegOpt.getOrElse(0), u.outDeg)
}.outerJoinVertices(initialUserGraph.outDegrees) {
    
    
    case (id, u, outDegOpt) => User(u.name, u.age, u.inDeg, outDegOpt.getOrElse(0))
}
//顶点的入度即为粉丝数量
for ((id, property) <- userGraph.vertices.collect) 
   println(s"User $id is ${property.name} and is liked by ${property.inDeg} people.")

• 结果如下

User 1 is called Alice and is liked by 2 people.
User 2 is called Bob and is liked by 2 people.
User 3 is called Charlie and is liked by 1 people.
User 4 is called David and is liked by 1 people.
User 5 is called Ed and is liked by 0 people.
User 6 is called Fran and is liked by 2 people.

6.练习一:谁是网络红人？

需求说明
数据：twitter-graph-data.txt
数据获取地址:链接: https://pan.baidu.com/s/1OM5DFo3qO_HDDmCS2PIaWQ 提取码: v8r3
格式：((User*, *),(User*,*))
(User*, *)=(用户名,用户ID)
第一个用户表示被跟随者（followee）
第二个用户表示跟随者（follower）
创建图并计算每个用户的粉丝数量
找出网络红人

import org.apache.spark.graphx._
//正则表达式截取数据
val pattern = """\(\((User[0-9]{1,},[0-9]{1,})\),\((User[0-9]{1,},[0-9]{1,})\)\)""".r
//加载数据
val t1 = sc.textFile("file:///data/twitter_graph_data.txt")
val t2 =  t1.map(line=>line match{
    
    case pattern(followee,follower)=>(Some(followee),Some(follower));case _ => (None,None)})
val t3 = t2.filter(x=>x._1!=None && x._2!=None).map(x=>(x._1.get.split(","),x._2.get.split(","))).map(x=>(x._1(0),x._1(1).toLong,x._2(0),x._2(1).toLong))
//形成顶点,边和图
val verts = t3.flatMap(x => Array((x._2, x._1), (x._4, x._3))).distinct
val edges = t3.map(x=>Edge(x._2,x._4,"follow"))
val graph = Graph(verts,edges)
val defaultUser = ("")
val graph = Graph(verts, edges, defaultUser)
//找出网络红人即粉丝数最大的人
graph.inDegrees.sortBy(_._2,false).take(1)
//找出影响力最大的红人
val sub_graph = graph.pregel("", 2, EdgeDirection.In)((_, attr, msg) => attr + "," + msg, triplet => Iterator((triplet.srcId, triplet.dstAttr)), (a, b) => (a + "," + b))

val lengthRDD = sub_graph.vertices.map(v=>(v._1,v._2.split(",").distinct.length-2)).max()(new Ordering[Tuple2[VertexId,Int]](){
    
    override def compare(x:(VertexId,Int),y:(VertexId,Int)):Int= Ordering[Int].compare(x._2,y._2)})

val userId = graph.vertices.filter(_._1 == lengthRDD._1).map(_._2).collect().head
println(userId + " has maximum influence on network with " + lengthRDD._2 + " influencers.")

7.PageRank in GraphX

PageRank（PR）算法

• 用于评估网页链接的质量和数量，以确定该网页的重要性和权威性的相对分数，范围为0到10
• 从本质上讲，PageRank是找出图中顶点（网页链接）的重要性
• GraphX提供了PageRank API用于计算图的PageRank

class Graph[VD, ED] {
    
    
//第一个参数为收敛时允许的误差，越小越精确， 确定迭代是否结束的参数
//第二个参数为随机重置概率
  def pageRank(tol: Double, resetProb: Double = 0.15): Graph[Double, Double]
}

PageRank应用

//找出用户社交网络中最重要的用户
val tweeters = Array((1L, ("Alice", 28)), (2L, ("Bob", 27)), (3L, ("Charlie", 65)), (4L, ("David", 42)), (5L, ("Ed", 55)), (6L, ("Fran", 50)))
val vertexRDD: RDD[(Long, (String, Int))] = spark.sparkContext.parallelize(tweeters)

val followRelations = Array(Edge[Int](2L, 1L, 7), Edge[Int](2L, 4L, 2), Edge[Int](3L, 2L, 4),  Edge[Int](3L, 6L, 3), Edge[Int](4L, 1L, 1), Edge[Int](5L, 2L, 2), Edge[Int](5L, 3L, 8), Edge[Int](5L, 6L, 3))
val edgeRDD = spark.sparkContext.parallelize(followRelations)

val graph: Graph[(String, Int), Int] = Graph(vertexRDD, edgeRDD)
val ranks = graph.pageRank(0.0001)
ranks.vertices.sortBy(_._2, false).collect

8.练习2：PageRank应用

/*需求说明
现有followers.txt、users.txt，通过followers.txt创建图，并使用PageRank算法找出图中最重要的用户，输出用户名称与重要程度
*/
//导包
import org.apache.spark.graphx.GraphLoader
//加载边数据形成图
val graph = GraphLoader.edgeListFile(sc, "file:///data/followers.txt")
//给定参数形成重要系数
val ranks = graph.pageRank(0.0001).vertices
//加载顶点数据
val users = sc.textFile("file:///data/user.txt").map {
    
     line =>
  val fields = line.split(",")
  (fields(0).toLong, fields(1))
}
//join操作形成名字和重要系数的rdd
val ranksByUsername = users.join(ranks).map {
    
    
  case (id, (username, rank)) => (username, rank)
}
//找出重要系数最大的人
ranksByUsername.sortBy(_._2,false).take(1)

9.Pregel概述

Pregel是Google提出的用于大规模分布式图计算框架

• 图遍历（BFS）
• 单源最短路径（SSSP）
• PageRank计算

Pregel的计算由一系列迭代组成，称为supersteps
Pregel迭代过程

• 每个顶点从上一个superstep接收入站消息
• 计算顶点新的属性值
• 在下一个superstep中向相邻的顶点发送消息
• 当没有剩余消息时，迭代结束

//源码
class Graph[VD, ED] {
    
      
    def pregel[A](initialMsg: A, maxIterations: Int, activeDirection: EdgeDirection)(
      vprog: (VertexID, VD, A) => VD,
      sendMsg: EdgeTriplet[VD, ED] => Iterator[(VertexID,A)],
      mergeMsg: (A, A) => A)
    : Graph[VD, ED]
}

//参数说明
initialMsg：在“superstep 0”之前发送至顶点的初始消息
maxIterations：将要执行的最大迭代次数
activeDirection：发送消息方向（默认是出边方向：EdgeDirection.Out）
vprog：用户定义函数，用于顶点接收消息
sendMsg：用户定义的函数，用于确定下一个迭代发送的消息及发往何处
mergeMsg：用户定义的函数，在vprog前，合并到达顶点的多个消息

GraphX Pregel 应用

//需求说明：求出图中最小值 
// 创建顶点集RDD
val vertices: RDD[(VertexId, (Int, Int))] = sc.parallelize(Array((1L, (7,-1)), (2L, (3,-1)),  (3L, (2,-1)), (4L, (6,-1))))
// 创建边集RDD
val relationships: RDD[Edge[Boolean]] = sc.parallelize(Array(Edge(1L, 2L, true), Edge(1L, 4L, true),  Edge(2L, 4L, true), Edge(3L, 1L, true),  Edge(3L, 4L, true)))
// 创建图
val graph = Graph(vertices, relationships)
//Pregel
val initialMsg = 9999//最大迭代次数
def vprog(vertexId: VertexId, value: (Int, Int), message: Int): (Int, Int) = {
    
    
  if (message == initialMsg)  value else (message min value._1, value._1)
}

def sendMsg(triplet: EdgeTriplet[(Int, Int), Boolean]): Iterator[(VertexId, Int)] = {
    
    
  val sourceVertex = triplet.srcAttr
  if (sourceVertex._1 == sourceVertex._2) Iterator.empty  else  Iterator((triplet.dstId, sourceVertex._1))
}

def mergeMsg(msg1: Int, msg2: Int): Int = msg1 min msg2
val minGraph = graph.pregel(initialMsg, Int.MaxValue,  EdgeDirection.Out)(vprog, sendMsg, mergeMsg)
minGraph.vertices.collect.foreach{
    
    
  case (vertexId, (value, original_value)) => println(value)
}

10.练习3：使用Pregel计算单源最短路径

/*需求说明
求从0到任意点的最短路径（SSSP）
实现思路
初始化 Vertex 的 Message 为最大值
将源点（0）的 Message 设为 0
每步每个节点将自己目前的 Message 加上边的权值发送到相邻节点，每个节点聚合出自身所有消息的最小值
当某一步当中一个节点Message 值无变化，该节点停止迭代
*/

• maven依赖 如下

    <dependencies>
        <dependency>
            <groupId>org.scala-lang</groupId>
            <artifactId>scala-library</artifactId>
            <version>2.11.8</version>
        </dependency>
        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-core_2.11</artifactId>
            <version>2.2.0</version>
        </dependency>
        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-sql_2.11</artifactId>
            <version>2.2.0</version>
        </dependency>
        <dependency>
            <groupId>org.apache.hadoop</groupId>
            <artifactId>hadoop-client</artifactId>
            <version>2.6.0</version>
        </dependency>
        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-graphx_2.11</artifactId>
            <version>2.2.0</version>
        </dependency>
    </dependencies>

object pregelDemo03 extends App{
    
    
  //求顶点0到各顶点最短距离
  val sc = new SparkContext( new SparkConf().setAppName("test").setMaster("local[*]"))
  val vertexRDD = sc.makeRDD(Array((0L,""),(1L,""),(2L,""),(3L,""),(4L,"")))
  val edgeRDD = sc.makeRDD(Array(Edge(0L,4L,10),Edge(4L,3L,50),Edge(0L,1L,100),Edge(0L,2L,30),Edge(3L,1L,10),Edge(2L,1L,60),Edge(2L,3L,60)))
  val graph =Graph(vertexRDD,edgeRDD)
  //起始顶点
  val srcVertextId = 0L
  val initialGraph = graph.mapVertices{
    
    case (id,property)=>if (id==srcVertextId) 0.0 else Double.PositiveInfinity}
  //调用pregel
  val pregelGraph = initialGraph.pregel(Double.PositiveInfinity,Int.MaxValue,EdgeDirection.Out)(
    (vid:VertexId,vd:Double,distMsg:Double)=>{
    
    
      val minDist = math.min(vd,distMsg)
      println(s"顶点${vid}，属性${vd}，收到消息${distMsg}，合并后的属性${minDist}")
      minDist
    },
    (edgeTriplet: EdgeTriplet[Double,PartitionID]) => {
    
    
      if (edgeTriplet.srcAttr + edgeTriplet.attr < edgeTriplet.dstAttr) {
    
    
        println(s"顶点${edgeTriplet.srcId} 给 顶点${edgeTriplet.dstId} 发送消息 ${edgeTriplet.srcAttr + edgeTriplet.attr}")
        Iterator[(VertexId, Double)]((edgeTriplet.dstId, edgeTriplet.srcAttr + edgeTriplet.attr))
      } else {
    
    
        Iterator.empty
      }
    },
    (msg1: Double, msg2: Double) => math.min(msg1, msg2)
  )
//  pregelGraph.triplets.collect().foreach(println)
    println(pregelGraph.vertices.collect.mkString("\n"))
    sc.stop()
}

• 运行结果如下