背景:
spark graphx并未提供极大团挖掘算法
当下的极大团算法都是串行化的算法,基于Bron–Kerbosch算法
思路:
spark graphx提供了连通图的算法,连通图和极大团都是无向图中的概念,极大团为连通图的子集
利用spark graphx 找出连通图,在从各个连通图中,利用串行化的极大团算法,找出极大团 (伪并行化)
对于关联性较强的图,找出来的连通图非常大,这时串行化的极大团算法,仍然会耗时很久,这里利用剪枝的思想减少样本数据量,但是对于大图,优化空间有限
期待真正的并行化的极大团算法
配置文件:
graph_data_path=hdfs://localhost/graph_data
out_path=hdfs://localhost/clique
ck_path=hdfs://localhost/checkpoint
numIter=50剪枝次数
count=3极大团顶点数大小
algorithm=2极大团算法,1:个人实现 2:jgrapht
percent=90剪枝后的顶点数,占前一次的百分比,如果剪完后,还剩下90%的数据,那么剪枝效率已然不高
spark.master=local
spark.app.name=graph
spark.serializer=org.apache.spark.serializer.KryoSerializer
spark.yarn.executor.memoryOverhead=20480
spark.yarn.driver.memoryOverhead=20480
spark.driver.extraJavaOptions=-XX:+UseG1GC -XX:+UseCompressedOops -XX:+DisableExplicitGC
spark.executor.extraJavaOptions=-XX:+UseG1GC -XX:+UseCompressedOops -XX:+DisableExplicitGC
spark.driver.maxResultSize=10g
spark.default.parallelism=60
样本数据:
{"src":"0","dst":"1"}
{"src":"0","dst":"2"}
{"src":"0","dst":"3"}
{"src":"1","dst":"0"}
{"src":"2","dst":"1"}
{"src":"3","dst":"5"}
{"src":"4","dst":"6"}
{"src":"5","dst":"4"}
{"src":"6","dst":"5"}
{"src":"3","dst":"2"}
{"src":"2","dst":"3"}
{"src":"6","dst":"4"}
{"src":"3","dst":"4"}
{"src":"4","dst":"3"}
{"src":"2","dst":"6"}
{"src":"6","dst":"2"}
{"src":"6","dst":"7"}
{"src":"7","dst":"6"}
样本图:
输出:
0,1,2
0,2,3
3,4,5
4,5,6
代码实现:
import java.util
import java.util.Properties
import org.apache.spark.broadcast.Broadcast
import org.apache.spark.graphx.{Edge, Graph}
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.{Row, SQLContext}
import org.apache.spark.storage.StorageLevel
import org.apache.spark.{SparkConf, SparkContext}
import org.jgrapht.alg.BronKerboschCliqueFinder
import org.jgrapht.graph.{DefaultEdge, SimpleGraph}
import scala.collection.JavaConverters._
import scala.collection.mutable
object ApplicationTitan {
def main(args: Array[String]) {
val prop = new Properties()
prop.load(getClass.getResourceAsStream("/config.properties"))
val graph_data_path = prop.getProperty("graph_data_path")
val out_path = prop.getProperty("out_path")
val ck_path = prop.getProperty("ck_path")
val count = Integer.parseInt(prop.getProperty("count"))
val numIter = Integer.parseInt(prop.getProperty("numIter"))
val algorithm = Integer.parseInt(prop.getProperty("algorithm"))
val percent = Integer.parseInt(prop.getProperty("percent"))
val conf = new SparkConf()
try {
Runtime.getRuntime.exec("hdfs dfs -rm -r " + out_path)
// Runtime.getRuntime.exec("cmd.exe /C rd /s /q " + out_path)
} catch {
case ex: Exception =>
ex.printStackTrace(System.out)
}
prop.stringPropertyNames().asScala.foreach(s => {
if (s.startsWith("spark")) {
conf.set(s, prop.getProperty(s))
}
})
conf.registerKryoClasses(Array(getClass))
val sc = new SparkContext(conf)
sc.setLogLevel("ERROR")
sc.setCheckpointDir(ck_path)
val sqlc = new SQLContext(sc)
try {
val e_df = sqlc.read
// .json(graph_data_path)
.parquet(graph_data_path)
var e_rdd = e_df
.mapPartitions(it => {
it.map({
case Row(dst: String, src: String) =>
val src_long = src.toLong
val dst_long = dst.toLong
if (src_long < dst_long) (src_long, dst_long) else (dst_long, src_long)
})
}).distinct()
e_rdd.persist(StorageLevel.MEMORY_AND_DISK_SER)
var bc: Broadcast[Set[Long]] = null
var iter = 0
var bc_size = 0
//剪枝
while (iter <= numIter) {
val temp = e_rdd
.flatMap(x => List((x._1, 1), (x._2, 1)))
.reduceByKey((x, y) => x + y)
.filter(x => x._2 >= count - 1)
.mapPartitions(it => it.map(x => x._1))
val bc_value = temp.collect().toSet
bc = sc.broadcast(bc_value)
e_rdd = e_rdd.filter(x => bc.value.contains(x._1) && bc.value.contains(x._2))
e_rdd.persist(StorageLevel.MEMORY_AND_DISK_SER)
iter += 1
if (bc_size != 0 && bc_value.size >= bc_size * percent / 100) {
println("total iter : "+ iter)
iter = Int.MaxValue
}
bc_size = bc_value.size
}
// 构造图
val edge: RDD[Edge[Long]] = e_rdd.mapPartitions(it => it.map(x => Edge(x._1, x._2)))
val graph = Graph.fromEdges(edge, 0, StorageLevel.MEMORY_AND_DISK_SER, StorageLevel.MEMORY_AND_DISK_SER)
//连通图
val cc = graph.connectedComponents().vertices
cc.persist(StorageLevel.MEMORY_AND_DISK_SER)
cc.join(e_rdd)
.mapPartitions(it => it.map(x => ((math.random * 10).toInt.toString.concat(x._2._1.toString), (x._1, x._2._2))))
.aggregateByKey(List[(Long, Long)]())((list, v) => list :+ v, (list1, list2) => list1 ::: list2)
.mapPartitions(it => it.map(x => (x._1.substring(1), x._2)))
.aggregateByKey(List[(Long, Long)]())((list1, list2) => list1 ::: list2, (list3, list4) => list3 ::: list4)
.filter(x => x._2.size >= count - 1)
.flatMap(x => {
if (algorithm == 1)
find(x, count)
else
find2(x, count)
})
.mapPartitions(it => {
it.map({
case set =>
var temp = ""
set.asScala.foreach(x => temp += x + ",")
temp.substring(0, temp.length - 1)
case _ =>
})
})
// .coalesce(1)
.saveAsTextFile(out_path)
}
catch {
case ex: Exception =>
ex.printStackTrace(System.out)
}
sc.stop()
}
//自己实现的极大团算法
def find(x: (String, List[(Long, Long)]), count: Int): mutable.Set[util.Set[String]] = {
println(x._1 + "|s|" + x._2.size)
println("BKCliqueFinder---" + x._1 + "---" + System.currentTimeMillis())
val neighbors = new util.HashMap[String, util.Set[String]]
val finder = new CliqueFinder(neighbors, count)
x._2.foreach(r => {
val v1 = r._1.toString
val v2 = r._2.toString
if (neighbors.containsKey(v1)) {
neighbors.get(v1).add(v2)
} else {
val temp = new util.HashSet[String]()
temp.add(v2)
neighbors.put(v1, temp)
}
if (neighbors.containsKey(v2)) {
neighbors.get(v2).add(v1)
} else {
val temp = new util.HashSet[String]()
temp.add(v1)
neighbors.put(v2, temp)
}
})
println("BKCliqueFinder---" + x._1 + "---" + System.currentTimeMillis())
finder.findMaxCliques().asScala
}
//jgrapht 中的极大团算法
def find2(x: (String, List[(Long, Long)]), count: Int): Set[util.Set[String]] = {
println(x._1 + "|s|" + x._2.size)
println("BKCliqueFinder---" + x._1 + "---" + System.currentTimeMillis())
val to_clique = new SimpleGraph[String, DefaultEdge](classOf[DefaultEdge])
x._2.foreach(r => {
val v1 = r._1.toString
val v2 = r._2.toString
to_clique.addVertex(v1)
to_clique.addVertex(v2)
to_clique.addEdge(v1, v2)
})
val finder = new BronKerboschCliqueFinder(to_clique)
val list = finder.getAllMaximalCliques.asScala
var result = Set[util.Set[String]]()
list.foreach(x => {
if (x.size() >= count)
result = result + x
})
println("BKCliqueFinder---" + x._1 + "---" + System.currentTimeMillis())
result
}
}
自己实现的极大团算法:
import java.util.*;
/**
* [@author](https://my.oschina.net/arthor) mopspecial@gmail.com
* [@date](https://my.oschina.net/u/2504391) 2017/7/31
*/
public class CliqueFinder {
private Map> neighbors;
private Set nodes;
private Set> maxCliques = new HashSet<>();
private Integer minSize;
public CliqueFinder(Map> neighbors, Integer minSize) {
this.neighbors = neighbors;
this.nodes = neighbors.keySet();
this.minSize = minSize;
}
private void bk3(Set clique, List candidates, List excluded) {
if (candidates.isEmpty() && excluded.isEmpty()) {
if (!clique.isEmpty() && clique.size() >= minSize) {
maxCliques.add(clique);
}
return;
}
for (String s : degeneracy_order(candidates)) {
List new_candidates = new ArrayList<>(candidates);
new_candidates.retainAll(neighbors.get(s));
List new_excluded = new ArrayList<>(excluded);
new_excluded.retainAll(neighbors.get(s));
Set nextClique = new HashSet<>(clique);
nextClique.add(s);
bk2(nextClique, new_candidates, new_excluded);
candidates.remove(s);
excluded.add(s);
}
}
private void bk2(Set clique, List candidates, List excluded) {
if (candidates.isEmpty() && excluded.isEmpty()) {
if (!clique.isEmpty() && clique.size() >= minSize) {
maxCliques.add(clique);
}
return;
}
String pivot = pick_random(candidates);
if (pivot == null) {
pivot = pick_random(excluded);
}
List tempc = new ArrayList<>(candidates);
tempc.removeAll(neighbors.get(pivot));
for (String s : tempc) {
List new_candidates = new ArrayList<>(candidates);
new_candidates.retainAll(neighbors.get(s));
List new_excluded = new ArrayList<>(excluded);
new_excluded.retainAll(neighbors.get(s));
Set nextClique = new HashSet<>(clique);
nextClique.add(s);
bk2(nextClique, new_candidates, new_excluded);
candidates.remove(s);
excluded.add(s);
}
}
private List degeneracy_order(List innerNodes) {
List result = new ArrayList<>();
Map deg = new HashMap<>();
for (String node : innerNodes) {
deg.put(node, neighbors.get(node).size());
}
while (!deg.isEmpty()) {
Integer min = Collections.min(deg.values());
String minKey = null;
for (String key : deg.keySet()) {
if (deg.get(key).equals(min)) {
minKey = key;
break;
}
}
result.add(minKey);
deg.remove(minKey);
for (String k : neighbors.get(minKey)) {
if (deg.containsKey(k)) {
deg.put(k, deg.get(k) - 1);
}
}
}
return result;
}
private String pick_random(List random) {
if (random != null && !random.isEmpty()) {
return random.get(0);
} else {
return null;
}
}
public Set> findMaxCliques() {
this.bk3(new HashSet<>(), new ArrayList<>(nodes), new ArrayList<>());
return maxCliques;
}
public static void main(String[] args) {
Map> neighbors = new HashMap<>();
neighbors.put("0", new HashSet<>(Arrays.asList("1", "2", "3")));
neighbors.put("1", new HashSet<>(Arrays.asList("0", "2")));
neighbors.put("2", new HashSet<>(Arrays.asList("0", "1", "3", "6")));
neighbors.put("3", new HashSet<>(Arrays.asList("0", "2", "4", "5")));
neighbors.put("4", new HashSet<>(Arrays.asList("3", "5", "6")));
neighbors.put("5", new HashSet<>(Arrays.asList("3", "4", "6")));
neighbors.put("6", new HashSet<>(Arrays.asList("2", "4", "5")));
neighbors.put("7", new HashSet<>(Arrays.asList("6")));
CliqueFinder finder = new CliqueFinder(neighbors, 3);
finder.bk3(new HashSet<>(), new ArrayList<>(neighbors.keySet()), new ArrayList<>());
System.out.println(finder.maxCliques);
}
}