使用Spark ALS實現協同過濾

轉自：http://blog.javachen.com/2015/06/01/how-to-implement-collaborative-filtering-using-spark-als.html

本文主要記錄最近一段時間學習和實現Spark MLlib中的協同過濾的一些總結，希望對大家熟悉Spark ALS算法有所幫助。

更新：

1. 【2016.06.12】Spark1.4.0中MatrixFactorizationModel提供了recommendForAll方法實現離線批量推薦，見SPARK-3066。

測試環境

爲了測試簡單，在本地以local方式運行Spark，你需要做的是下載編譯好的壓縮包解壓即可，可以參考Spark本地模式運行。

測試數據使用MovieLens的MovieLens 10M數據集，下載之後解壓到data目錄。數據的格式請參考README中的說明，需要注意的是ratings.dat中的數據被處理過，每個用戶至少訪問了20個商品。

下面的代碼均在spark-shell中運行，啓動時候可以根據你的機器內存設置JVM參數，例如：

bin/spark-shell --executor-memory 3g --driver-memory 3g --driver-java-options '-Xms2g -Xmx2g -XX:+UseCompressedOops'

預測評分

這個例子主要演示如何訓練數據、評分並計算根均方差。

準備工作

首先，啓動spark-shell，然後引入mllib包，我們需要用到ALS算法類和Rating評分類：

import org.apache.spark.mllib.recommendation.{ALS, Rating}

Spark的日誌級別默認爲INFO，你可以手動設置爲WARN級別，同樣先引入log4j依賴：

import org.apache.log4j.{Logger,Level}

然後，運行下面代碼：

Logger.getLogger("org.apache.spark").setLevel(Level.WARN)
Logger.getLogger("org.eclipse.jetty.server").setLevel(Level.OFF)

加載數據

spark-shell啓動成功之後，sc爲內置變量，你可以通過它來加載測試數據：

val data = sc.textFile("data/ml-1m/ratings.dat")

接下來解析文件內容，獲得用戶對商品的評分記錄：

val ratings = data.map(_.split("::") match { case Array(user, item, rate, ts) =>
  Rating(user.toInt, item.toInt, rate.toDouble)
}).cache()

查看第一條記錄：

scala> ratings.first
res81: org.apache.spark.mllib.recommendation.Rating = Rating(1,1193,5.0)

我們可以統計文件中用戶和商品數量：

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.")

可以看到如下輸出：

//Got 1000209 ratings from 6040 users on 3706 products.

你可以對評分數據生成訓練集和測試集，例如：訓練集和測試集比例爲8比2：

val splits = ratings.randomSplit(Array(0.8, 0.2), seed = 111l)
val training = splits(0).repartition(numPartitions)
val test = splits(1).repartition(numPartitions)

這裏，我們是將評分數據全部當做訓練集，並且也爲測試集。

訓練模型

接下來調用ALS.train()方法，進行模型訓練：

val rank = 12
val lambda = 0.01
val numIterations = 20
val model = ALS.train(ratings, rank, numIterations, lambda)

訓練完後，我們看看model中的用戶和商品特徵向量：

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

評測

我們要對比一下預測的結果，注意：我們將訓練集當作測試集來進行對比測試。從訓練集中獲取用戶和商品的映射：

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

顯然，測試集的記錄數等於評分總記錄數，驗證一下：

usersProducts.count  //Long = 1000209

使用推薦模型對用戶商品進行預測評分，得到預測評分的數據集：

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

查看其記錄數：

predictions.count //Long = 1000209

將真實評分數據集與預測評分數據集進行合併，這樣得到用戶對每一個商品的實際評分和預測評分：

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

ratesAndPreds.count  //Long = 1000209

然後計算根均方差：

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

println(s"RMSE = $rmse")

上面這段代碼其實就是對測試集進行評分預測並計算相似度，這段代碼可以抽象爲一個方法，如下：

/** 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())
}

除了RMSE指標，我們還可以及時AUC以及Mean average precision at K (MAPK)，關於AUC的計算方法，參考RunRecommender.scala，關於MAPK的計算方法可以參考《Packt.Machine Learning with Spark.2015.pdf》一書第四章節內容，或者你可以看本文後面內容。

保存真實評分和預測評分

我們還可以保存用戶對商品的真實評分和預測評分記錄到本地文件：

ratesAndPreds.sortByKey().repartition(1).sortBy(_._1).map({
  case ((user, product), (rate, pred)) => (user + "," + product + "," + rate + "," + pred)
}).saveAsTextFile("/tmp/result")

上面這段代碼先按用戶排序，然後重新分區確保目標目錄中只生成一個文件。如果你重複運行這段代碼，則需要先刪除目標路徑：

import scala.sys.process._
"rm -r /tmp/result".!

我們還可以對預測的評分結果按用戶進行分組並按評分倒排序：

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})
}

給一個用戶推薦商品

這個例子主要是記錄如何給一個或大量用戶進行推薦商品，例如，對用戶編號爲384的用戶進行推薦，查出該用戶在測試集中評分過的商品。

找出5個用戶：

users.take(5) 
//Array[Int] = Array(384, 1084, 4904, 3702, 5618)

查看用戶編號爲384的用戶的預測結果中預測評分排前10的商品：

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)

查看該用戶的評分記錄：

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)

可以看到該用戶對22個商品評過分以及瀏覽的商品是哪些。

我們可以該用戶對某一個商品的實際評分和預測評分方差爲多少：

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

如何找出和一個已知商品最相似的商品呢？這裏，我們可以使用餘弦相似度來計算：

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())
}

以2055商品爲例，計算實際評分和預測評分相似度

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

找到和該商品最相似的10個商品：

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)

顯然第一個最相似的商品即爲該商品本身，即2055，我們可以修改下代碼，取前k+1個商品，然後排除第一個：

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)

接下來，我們可以計算給該用戶推薦的前K個商品的平均準確度MAPK，該算法定義如下（該算法是否正確還有待考證）：

/* 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
  }
}

給該用戶推薦的商品爲：

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)

給該用戶預測的商品爲：

 val predictedProducts = topKRecs.map(_.product)
//predictedProducts: Array[Int] = Array(2545, 129, 184, 811, 1421, 1313, 2892, 2483, 397, 97)

最後的準確度爲：

val apk10 = avgPrecisionK(actualProducts, predictedProducts, 10)
// apk10: Double = 0.0

批量推薦

你可以評分記錄中獲得所有用戶然後依次給每個用戶推薦：

val users = ratings.map(_.user).distinct()

users.collect.flatMap { user =>
  model.recommendProducts(user, 10)
}

這種方式是遍歷內存中的一個集合然後循環調用RDD的操作，運行會比較慢，另外一種方式是直接操作model中的userFeatures和productFeatures，代碼如下：

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)

//獲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)
  //根據索引取對應的商品id
  val recommendedProducts = sortedWithId.map(_._2).map{idx=>idxProductsBroadcast.value.get(idx).get}
  (user, recommendedProducts) 
}

這種方式其實還不是最優方法，更好的方法可以參考Personalised recommendations using Spark，當然這篇文章中的代碼還可以繼續優化一下。我修改後的代碼如下，供大家參考：

val productFeatures = model.productFeatures.collect()
var productArray = ArrayBuffer[Int]()
var productFeaturesArray = ArrayBuffer[Array[Double]]()
for ((product, features) <- productFeatures) {
  productArray += product
  productFeaturesArray += features
}

val productArrayBroadcast = sc.broadcast(productArray)
val productFeatureMatrixBroadcast = sc.broadcast(new DoubleMatrix(productFeaturesArray.toArray).transpose())

start = System.currentTimeMillis()
val allRecs = model.userFeatures.mapPartitions { iter =>
  // Build user feature matrix for jblas
  var userFeaturesArray = ArrayBuffer[Array[Double]]()
  var userArray = new ArrayBuffer[Int]()
  while (iter.hasNext) {
    val (user, features) = iter.next()
    userArray += user
    userFeaturesArray += features
  }

  var userFeatureMatrix = new DoubleMatrix(userFeaturesArray.toArray)
  var userRecommendationMatrix = userFeatureMatrix.mmul(productFeatureMatrixBroadcast.value)
  var productArray=productArrayBroadcast.value
  var mappedUserRecommendationArray = new ArrayBuffer[String](params.topk)

  // Extract ratings from the matrix
  for (i <- 0 until userArray.length) {
    var ratingSet =  mutable.TreeSet.empty(Ordering.fromLessThan[(Int,Double)](_._2 > _._2))
    for (j <- 0 until productArray.length) {
      var rating = (productArray(j), userRecommendationMatrix.get(i,j))
      ratingSet += rating
    }
    mappedUserRecommendationArray += userArray(i)+","+ratingSet.take(params.topk).mkString(",")
  }
  mappedUserRecommendationArray.iterator
}

2015.06.12 更新：

悲哀的是，上面的方法還是不能解決問題，因爲矩陣相乘會撐爆集羣內存；可喜的是，如果你關注Spark最新動態，你會發現Spark1.4.0中MatrixFactorizationModel提供了recommendForAll方法實現離線批量推薦，詳細說明見SPARK-3066。因爲，我使用的Hadoop版本是CDH-5.4.0，其中Spark版本還是1.3.0，所以暫且不能在集羣上測試Spark1.4.0中添加的新方法。

如果上面結果跑出來了，就可以驗證推薦結果是否正確。還是以384用戶爲例：

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)

接下來，我們可以計算所有推薦結果的準確度了，首先，得到每個用戶評分過的所有商品：

val userProducts = ratings.map{ case Rating(user, product, rating) => (user, product) }.groupBy(_._1)

然後，預測的商品和實際商品關聯求準確度：

// 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

其實，我們也可以使用Spark內置的算法計算RMSE和MAE：

// 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

計算推薦2000個商品時的準確度爲：

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

保存和加載推薦模型

對與實時推薦，我們需要啓動一個web server，在啓動的時候生成或加載訓練模型，然後提供API接口返回推薦接口，需要調用的相關方法爲：

save(model: MatrixFactorizationModel, path: String)
load(sc: SparkContext, path: String)

model中的userFeatures和productFeatures也可以保存起來：

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")

總結

本文主要記錄如何使用ALS算法實現協同過濾並給用戶推薦商品，以上代碼在Github倉庫中的ScalaLocalALS.scala文件。

如果你想更加深入瞭解Spark MLlib算法的使用，可以看看Packt.Machine Learning with Spark.2015.pdf這本電子書並下載書中的源碼，本文大部分代碼參考自該電子書。