分享

如何使用Spark ALS实现协同过滤

 
lovebyz 2015-6-9 17:06:57 发表于 实操演练 [显示全部楼层] 回帖奖励 阅读模式 关闭右栏 60 110433




测试环境
为了测试简单,在本地以local方式运行Spark,你需要做的是下载编译好的压缩包解压即可,可以参考[url=Spark本地运行模式 http://www.aboutyun.com/thread-13618-1-1.html (出处: about云开发)]Spark本地模式运行[/url]。
测试数据使用MovieLensMovieLens 10M数据集,下载之后解压到data目录。数据的格式请参考README中的说明,需要注意的是ratings.dat中的数据被处理过,每个用户至少访问了20个商品。
下面的代码均在spark-shell中运行,启动时候可以根据你的机器内存设置JVM参数,例如:
[mw_shl_code=bash,true]bin/spark-shell --executor-memory 3g --driver-memory 3g --driver-java-options '-Xms2g -Xmx2g -XX:+UseCompressedOops'[/mw_shl_code]

预测评分
这个例子主要演示如何训练数据、评分并计算根均方差。
准备工作首先,启动spark-shell,然后引入mllib包,我们需要用到ALS算法类和Rating评分类:[mw_shl_code=scala,true]import org.apache.spark.mllib.recommendation.{ALS, Rating}[/mw_shl_code]
Spark的日志级别默认为INFO,你可以手动设置为WARN级别,同样先引入log4j依赖:
[mw_shl_code=scala,true]import org.apache.log4j.{Logger,Level}
[/mw_shl_code]
然后,运行下面代码:
[mw_shl_code=scala,true]Logger.getLogger("org.apache.spark").setLevel(Level.WARN)
Logger.getLogger("org.eclipse.jetty.server").setLevel(Level.OFF)[/mw_shl_code]

加载数据spark-shell启动成功之后,sc为内置变量,你可以通过它来加载测试数据:
[mw_shl_code=scala,true]val data = sc.textFile("data/ml-1m/ratings.dat")
[/mw_shl_code]
接下来解析文件内容,获得用户对商品的评分记录:
[mw_shl_code=scala,true]val ratings = data.map(_.split("::") match { case Array(user, item, rate, ts) =>  Rating(user.toInt, item.toInt, rate.toDouble)
}).cache()
[/mw_shl_code]

查看第一条记录:
[mw_shl_code=scala,true]scala> ratings.first
res81: org.apache.spark.mllib.recommendation.Rating = Rating(1,1193,5.0)
[/mw_shl_code]

我们可以统计文件中用户和商品数量:
[mw_shl_code=scala,true]val users = ratings.map(_.user).distinct()
val products = ratings.map(_.product).distinct()
println("Got "+ratings.count()+" ratings from "+users.count+" users on "+products.count+" products.")
[/mw_shl_code]
可以看到如下输出:
[mw_shl_code=scala,false]//Got 1000209 ratings from 6040 users on 3706 products.[/mw_shl_code]

你可以对评分数据生成训练集和测试集,例如:训练集和测试集比例为8比2:
[mw_shl_code=scala,true]val splits = ratings.randomSplit(Array(0.8, 0.2), seed = 111l)
val training = splits(0).repartition(numPartitions)
val test = splits(1).repartition(numPartitions)
[/mw_shl_code]
这里,我们是将评分数据全部当做训练集,并且也为测试集。


训练模型接下来调用ALS.train()方法,进行模型训练:
[mw_shl_code=scala,true]val rank = 12
val lambda = 0.01
val numIterations = 20
val model = ALS.train(ratings, rank, numIterations, lambda)[/mw_shl_code]

训练完后,我们看看model中的用户和商品特征向量:

[mw_shl_code=scala,false]model.userFeatures
//res82: org.apache.spark.rdd.RDD[(Int, Array[Double])] = users MapPartitionsRDD[400] at mapValues at ALS.scala:218

model.userFeatures.count
//res84: Long = 6040

model.productFeatures
//res85: org.apache.spark.rdd.RDD[(Int, Array[Double])] = products MapPartitionsRDD[401] at mapValues at ALS.scala:222

model.productFeatures.count
//res86: Long = 3706
[/mw_shl_code]

评测我们要对比一下预测的结果,注意:我们将训练集当作测试集来进行对比测试。从训练集中获取用户和商品的映射:
[mw_shl_code=scala,false]val usersProducts= ratings.map { case Rating(user, product, rate) =>
  (user, product)
}
[/mw_shl_code]

显然,测试集的记录数等于评分总记录数,验证一下:
[mw_shl_code=scala,false]usersProducts.count  //Long = 1000209
[/mw_shl_code]

使用推荐模型对用户商品进行预测评分,得到预测评分的数据集:
[mw_shl_code=scala,false]var predictions = model.predict(usersProducts).map { case Rating(user, product, rate) =>
    ((user, product), rate)
}
[/mw_shl_code]

查看其记录数:
[mw_shl_code=scala,false]predictions.count //Long = 1000209
[/mw_shl_code]

将真实评分数据集与预测评分数据集进行合并,这样得到用户对每一个商品的实际评分和预测评分:
[mw_shl_code=scala,false]val ratesAndPreds = ratings.map { case Rating(user, product, rate) =>
  ((user, product), rate)
}.join(predictions)

ratesAndPreds.count  //Long = 1000209[/mw_shl_code]

然后计算根均方差:
[mw_shl_code=scala,false]val rmse= math.sqrt(ratesAndPreds.map { case ((user, product), (r1, r2)) =>
  val err = (r1 - r2)
  err * err
}.mean())

println(s"RMSE = $rmse")[/mw_shl_code]

上面这段代码其实就是对测试集进行评分预测并计算相似度,这段代码可以抽象为一个方法,如下:
[mw_shl_code=scala,true]/** Compute RMSE (Root Mean Squared Error). */
def computeRmse(model: MatrixFactorizationModel, data: RDD[Rating]) = {
  val usersProducts = data.map { case Rating(user, product, rate) =>
    (user, product)
  }

  val predictions = model.predict(usersProducts).map { case Rating(user, product, rate) =>
    ((user, product), rate)
  }

  val ratesAndPreds = data.map { case Rating(user, product, rate) =>
    ((user, product), rate)
  }.join(predictions)

  math.sqrt(ratesAndPreds.map { case ((user, product), (r1, r2)) =>
    val err = (r1 - r2)
    err * err
  }.mean())
}[/mw_shl_code]

除了RMSE指标,我们还可以及时AUC以及Mean average precision at K (MAPK),关于AUC的计算方法,参考RunRecommender.scala,关于MAPK的计算方法可以参考《Packt.Machine Learning with Spark.2015.pdf》一书第四章节内容,或者你可以看本文后面内容。


保存真实评分和预测评分我们还可以保存用户对商品的真实评分和预测评分记录到本地文件:
[mw_shl_code=scala,false]ratesAndPreds.sortByKey().repartition(1).sortBy(_._1).map({
  case ((user, product), (rate, pred)) => (user + "," + product + "," + rate + "," + pred)
}).saveAsTextFile("/tmp/result")
[/mw_shl_code]

上面这段代码先按用户排序,然后重新分区确保目标目录中只生成一个文件。如果你重复运行这段代码,则需要先删除目标路径:
[mw_shl_code=scala,false]import scala.sys.process._
"rm -r /tmp/result".![/mw_shl_code]

我们还可以对预测的评分结果按用户进行分组并按评分倒排序:
[mw_shl_code=scala,false]predictions.map { case ((user, product), rate) =>
  (user, (product, rate))
}.groupByKey(numPartitions).map{case (user_id,list)=>
  (user_id,list.toList.sortBy {case (goods_id,rate)=> - rate})
}
[/mw_shl_code]



给一个用户推荐商品

这个例子主要是记录如何给一个或大量用户进行推荐商品,例如,对用户编号为384的用户进行推荐,查出该用户在测试集中评分过的商品。

找出5个用户:
[mw_shl_code=scala,false]users.take(5)
//Array[Int] = Array(384, 1084, 4904, 3702, 5618)[/mw_shl_code]

查看用户编号为384的用户的预测结果中预测评分排前10的商品:
[mw_shl_code=scala,true]val userId = users.take(1)(0) //384
val K = 10
val topKRecs = model.recommendProducts(userId, K)
println(topKRecs.mkString("\n"))
//    Rating(384,2545,8.354966018818265)
//    Rating(384,129,8.113083736094676)
//    Rating(384,184,8.038113395650853)
//    Rating(384,811,7.983433591425284)
//    Rating(384,1421,7.912044967873945)
//    Rating(384,1313,7.719639594879865)
//    Rating(384,2892,7.53667094600392)
//    Rating(384,2483,7.295378004543803)
//    Rating(384,397,7.141158013610967)
//    Rating(384,97,7.071089782695754)[/mw_shl_code]

查看该用户的评分记录:
[mw_shl_code=scala,true]val goodsForUser=ratings.keyBy(_.user).lookup(384)
// Seq[org.apache.spark.mllib.recommendation.Rating] = WrappedArray(Rating(384,2055,2.0), Rating(384,1197,4.0), Rating(384,593,5.0), Rating(384,599,3.0), Rating(384,673,2.0), Rating(384,3037,4.0), Rating(384,1381,2.0), Rating(384,1610,4.0), Rating(384,3074,4.0), Rating(384,204,4.0), Rating(384,3508,3.0), Rating(384,1007,3.0), Rating(384,260,4.0), Rating(384,3487,3.0), Rating(384,3494,3.0), Rating(384,1201,5.0), Rating(384,3671,5.0), Rating(384,1207,4.0), Rating(384,2947,4.0), Rating(384,2951,4.0), Rating(384,2896,2.0), Rating(384,1304,5.0))

productsForUser.size //Int = 22
productsForUser.sortBy(-_.rating).take(10).map(rating => (rating.product, rating.rating)).foreach(println)
//    (593,5.0)
//    (1201,5.0)
//    (3671,5.0)
//    (1304,5.0)
//    (1197,4.0)
//    (3037,4.0)
//    (1610,4.0)
//    (3074,4.0)
//    (204,4.0)
//    (260,4.0)
[/mw_shl_code]

可以看到该用户对22个商品评过分以及浏览的商品是哪些。

我们可以该用户对某一个商品的实际评分和预测评分方差为多少:
[mw_shl_code=scala,true]val actualRating = productsForUser.take(1)(0)
//actualRating: org.apache.spark.mllib.recommendation.Rating = Rating(384,2055,2.0)    val predictedRating = model.predict(789, actualRating.product)
val predictedRating = model.predict(384, actualRating.product)
//predictedRating: Double = 1.9426030777174637
val squaredError = math.pow(predictedRating - actualRating.rating, 2.0)
//squaredError: Double = 0.0032944066875075172[/mw_shl_code]

如何找出和一个已知商品最相似的商品呢?这里,我们可以使用余弦相似度来计算:
[mw_shl_code=scala,false]import org.jblas.DoubleMatrix

/* Compute the cosine similarity between two vectors */
def cosineSimilarity(vec1: DoubleMatrix, vec2: DoubleMatrix): Double = {
  vec1.dot(vec2) / (vec1.norm2() * vec2.norm2())
}
[/mw_shl_code]



以2055商品为例,计算实际评分和预测评分相似度
[mw_shl_code=scala,true]val itemId = 2055
val itemFactor = model.productFeatures.lookup(itemId).head
//itemFactor: Array[Double] = Array(0.3660752773284912, 0.43573060631752014, -0.3421429991722107, 0.44382765889167786, -1.4875195026397705, 0.6274569630622864, -0.3264533579349518, -0.9939845204353333, -0.8710321187973022, -0.7578890323638916, -0.14621856808662415, -0.7254264950752258)
val itemVector = new DoubleMatrix(itemFactor)
//itemVector: org.jblas.DoubleMatrix = [0.366075; 0.435731; -0.342143; 0.443828; -1.487520; 0.627457; -0.326453; -0.993985; -0.871032; -0.757889; -0.146219; -0.725426]

cosineSimilarity(itemVector, itemVector)
// res99: Double = 0.9999999999999999
[/mw_shl_code]

找到和该商品最相似的10个商品:
[mw_shl_code=scala,true]val sims = model.productFeatures.map{ case (id, factor) =>
  val factorVector = new DoubleMatrix(factor)
  val sim = cosineSimilarity(factorVector, itemVector)
  (id, sim)
}
val sortedSims = sims.top(K)(Ordering.by[(Int, Double), Double] { case (id, similarity) => similarity })
//sortedSims: Array[(Int, Double)] = Array((2055,0.9999999999999999), (2051,0.9138311231145874), (3520,0.8739823400539756), (2190,0.8718466671129721), (2050,0.8612639515847019), (1011,0.8466911667526461), (2903,0.8455764332511272), (3121,0.8227325520485377), (3674,0.8075743004357392), (2016,0.8063817280259447))
println(sortedSims.mkString("\n"))
//    (2055,0.9999999999999999)
//    (2051,0.9138311231145874)
//    (3520,0.8739823400539756)
//    (2190,0.8718466671129721)
//    (2050,0.8612639515847019)
//    (1011,0.8466911667526461)
//    (2903,0.8455764332511272)
//    (3121,0.8227325520485377)
//    (3674,0.8075743004357392)
//    (2016,0.8063817280259447)[/mw_shl_code]


显然第一个最相似的商品即为该商品本身,即2055,我们可以修改下代码,取前k+1个商品,然后排除第一个:
[mw_shl_code=scala,true]val sortedSims2 = sims.top(K + 1)(Ordering.by[(Int, Double), Double] { case (id, similarity) => similarity })
//sortedSims2: Array[(Int, Double)] = Array((2055,0.9999999999999999), (2051,0.9138311231145874), (3520,0.8739823400539756), (2190,0.8718466671129721), (2050,0.8612639515847019), (1011,0.8466911667526461), (2903,0.8455764332511272), (3121,0.8227325520485377), (3674,0.8075743004357392), (2016,0.8063817280259447), (3672,0.8016276723120674))

sortedSims2.slice(1, 11).map{ case (id, sim) => (id, sim) }.mkString("\n")
//    (2051,0.9138311231145874)
//    (3520,0.8739823400539756)
//    (2190,0.8718466671129721)
//    (2050,0.8612639515847019)
//    (1011,0.8466911667526461)
//    (2903,0.8455764332511272)
//    (3121,0.8227325520485377)
//    (3674,0.8075743004357392)
//    (2016,0.8063817280259447)
//    (3672,0.8016276723120674)[/mw_shl_code]

接下来,我们可以计算给该用户推荐的前K个商品的平均准确度MAPK,该算法定义如下(该算法是否正确还有待考证):
[mw_shl_code=scala,true]/* Function to compute average precision given a set of actual and predicted ratings */
// Code for this function is based on: https://github.com/benhamner/Metrics
def avgPrecisionK(actual: Seq[Int], predicted: Seq[Int], k: Int): Double = {
  val predK = predicted.take(k)
  var score = 0.0
  var numHits = 0.0
  for ((p, i) <- predK.zipWithIndex) {
    if (actual.contains(p)) {
      numHits += 1.0
      score += numHits / (i.toDouble + 1.0)
    }
  }
  if (actual.isEmpty) {
    1.0
  } else {
    score / scala.math.min(actual.size, k).toDouble
  }
}[/mw_shl_code]

给该用户推荐的商品为:

[mw_shl_code=scala,false]val actualProducts = productsForUser.map(_.product)
//actualProducts: Seq[Int] = ArrayBuffer(2055, 1197, 593, 599, 673, 3037, 1381, 1610, 3074, 204, 3508, 1007, 260, 3487, 3494, 1201, 3671, 1207, 2947, 2951, 2896, 1304)[/mw_shl_code]
给该用户预测的商品为:
[mw_shl_code=scala,false] val predictedProducts = topKRecs.map(_.product)
//predictedProducts: Array[Int] = Array(2545, 129, 184, 811, 1421, 1313, 2892, 2483, 397, 97)[/mw_shl_code]

最后的准确度为:
[mw_shl_code=scala,false]val apk10 = avgPrecisionK(actualProducts, predictedProducts, 10)
// apk10: Double = 0.0[/mw_shl_code]


批量推荐

你可以评分记录中获得所有用户然后依次给每个用户推荐:
[mw_shl_code=scala,false]val users = ratings.map(_.user).distinct()

users.collect.flatMap { user =>
  model.recommendProducts(user, 10)
}[/mw_shl_code]

这种方式是遍历内存中的一个集合然后循环调用RDD的操作,运行会比较慢,另外一种方式是直接操作model中的userFeatures和productFeatures,代码如下:
[mw_shl_code=scala,true]val itemFactors = model.productFeatures.map { case (id, factor) => factor }.collect()
val itemMatrix = new DoubleMatrix(itemFactors)
println(itemMatrix.rows, itemMatrix.columns)
//(3706,12)

// broadcast the item factor matrix
val imBroadcast = sc.broadcast(itemMatrix)

//获取商品和索引的映射
var idxProducts=model.productFeatures.map { case (prodcut, factor) => prodcut }.zipWithIndex().map{case (prodcut, idx) => (idx,prodcut)}.collectAsMap()
val idxProductsBroadcast = sc.broadcast(idxProducts)

val allRecs = model.userFeatures.map{ case (user, array) =>
  val userVector = new DoubleMatrix(array)
  val scores = imBroadcast.value.mmul(userVector)
  val sortedWithId = scores.data.zipWithIndex.sortBy(-_._1)
  //根据索引取对应的商品id
  val recommendedProducts = sortedWithId.map(_._2).map{idx=>idxProductsBroadcast.value.get(idx).get}
  (user, recommendedProducts)
}[/mw_shl_code]


这种方式其实还不是最优方法,更好的方法可以参考Personalised recommendations using Spark,当然这篇文章中的代码还可以继续优化一下。
验证推荐结果是否正确,还是以384用户为例:
[mw_shl_code=scala,true]allRecs.lookup(384).head.take(10)
//res50: Array[Int] = Array(1539, 219, 1520, 775, 3161, 2711, 2503, 771, 853, 759)
topKRecs.map(_.product)
//res49: Array[Int] = Array(1539, 219, 1520, 775, 3161, 2711, 2503, 771, 853, 759)[/mw_shl_code]

接下来,我们可以计算所有推荐结果的准确度了,首先,得到每个用户评分过的所有商品:
[mw_shl_code=scala,true]val userProducts = ratings.map{ case Rating(user, product, rating) => (user, product) }.groupBy(_._1)[/mw_shl_code]

然后,预测的商品和实际商品关联求准确度:
[mw_shl_code=scala,true]// finally, compute the APK for each user, and average them to find MAPK
val MAPK = allRecs.join(userProducts).map{ case (userId, (predicted, actualWithIds)) =>
  val actual = actualWithIds.map(_._2).toSeq
  avgPrecisionK(actual, predicted, K)
}.reduce(_ + _) / allRecs.count
println("Mean Average Precision at K = " + MAPK)
//Mean Average Precision at K = 0.018827551771260383[/mw_shl_code]

其实,我们也可以使用Spark内置的算法计算RMSE和MAE:
[mw_shl_code=scala,true]// MSE, RMSE and MAE
import org.apache.spark.mllib.evaluation.RegressionMetrics

val predictedAndTrue = ratesAndPreds.map { case ((user, product), (actual, predicted)) => (actual, predicted) }
val regressionMetrics = new RegressionMetrics(predictedAndTrue)
println("Mean Squared Error = " + regressionMetrics.meanSquaredError)
println("Root Mean Squared Error = " + regressionMetrics.rootMeanSquaredError)
// Mean Squared Error = 0.5490153087908566
// Root Mean Squared Error = 0.7409556726220918

// MAPK
import org.apache.spark.mllib.evaluation.RankingMetrics
val predictedAndTrueForRanking = allRecs.join(userProducts).map{ case (userId, (predicted, actualWithIds)) =>
  val actual = actualWithIds.map(_._2)
  (predicted.toArray, actual.toArray)
}
val rankingMetrics = new RankingMetrics(predictedAndTrueForRanking)
println("Mean Average Precision = " + rankingMetrics.meanAveragePrecision)
// Mean Average Precision = 0.04417535679520426[/mw_shl_code]

计算推荐2000个商品时的准确度为:
[mw_shl_code=scala,true]val MAPK2000 = allRecs.join(userProducts).map{ case (userId, (predicted, actualWithIds)) =>
  val actual = actualWithIds.map(_._2).toSeq
  avgPrecisionK(actual, predicted, 2000)
}.reduce(_ + _) / allRecs.count
println("Mean Average Precision = " + MAPK2000)
//Mean Average Precision = 0.025228311843069083[/mw_shl_code]


保存和加载推荐模型

对与实时推荐,我们需要启动一个web server,在启动的时候生成或加载训练模型,然后提供API接口返回推荐接口,需要调用的相关方法为:
[mw_shl_code=scala,false]save(model: MatrixFactorizationModel, path: String)
load(sc: SparkContext, path: String)[/mw_shl_code]

model中的userFeatures和productFeatures也可以保存起来:
[mw_shl_code=scala,true]val outputDir="/tmp"
model.userFeatures.map{ case (id, vec) => id + "\t" + vec.mkString(",") }.saveAsTextFile(outputDir + "/userFeatures")
model.productFeatures.map{ case (id, vec) => id + "\t" + vec.mkString(",") }.saveAsTextFile(outputDir + "/productFeatures")[/mw_shl_code]
总结
本文主要记录如何使用ALS算法实现协同过滤并给用户推荐商品,以上代码在Github仓库中的ScalaLocalALS.scala文件。
如果你想更加深入了解Spark MLlib算法的使用,可以看看Packt.Machine Learning with Spark.2015.pdf这本电子书并下载书中的源码,本文大部分代码参考自该电子书。
回复查看密码:
游客,如果您要查看本帖隐藏内容请回复



本文主要记录最近一段时间学习和实现Spark MLlib中的协同过滤的一些总结,
                 希望对大家熟悉Spark ALS算法有所帮助。


相关帖子

已有(60)人评论

跳转到指定楼层
zhangyi_bac 发表于 2015-6-10 07:28:19
回复

使用道具 举报

小南3707 发表于 2015-6-10 09:19:05
赞!         
回复

使用道具 举报

diandidemeng 发表于 2015-6-10 14:58:20
感谢分享……谢谢!
回复

使用道具 举报

6111111 发表于 2015-7-17 15:51:06
很有价值的资料,谢谢
回复

使用道具 举报

wanna 发表于 2015-7-23 13:00:35
感谢楼主分享
回复

使用道具 举报

zhangblh 发表于 2015-11-5 18:08:16
多谢楼主分享!!!棒
回复

使用道具 举报

xpy888 发表于 2015-12-2 18:01:08
很不错的文章
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

关闭

推荐上一条 /2 下一条