学习进度笔记

学习进度笔记29

协同过滤

import java.io.File

import scala.io.Source

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

import org.apache.spark.SparkConf

import org.apache.spark.SparkContext

import org.apache.spark.SparkContext._

import org.apache.spark.rdd._

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

object MovieLensALS {

  def main(args: Array[String]) {

    // 屏蔽不必要的日志显示在终端上

    Logger.getLogger("org.apache.spark").setLevel(Level.WARN)

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

    if (args.length != 2) {

      println("Usage: /path/to/spark/bin/spark-submit --driver-memory 2g --class week7.MovieLensALS " +

        "week7.jar movieLensHomeDir personalRatingsFile")

      sys.exit(1)

    }

    // 设置运行环境

    val conf = new SparkConf().setAppName("MovieLensALS").setMaster("local[4]")

    val sc = new SparkContext(conf)

    // 装载用户评分，该评分由评分器生成

    val myRatings = loadRatings(args(1))

    val myRatingsRDD = sc.parallelize(myRatings, 1)

    // 样本数据目录

    val movieLensHomeDir = args(0)

    // 装载样本评分数据，其中最后一列Timestamp取除10的余数作为key，Rating为值,即(Int,Rating)

    val ratings = sc.textFile(new File(movieLensHomeDir, "ratings.dat").toString).map { line =>

      val fields = line.split("::")

      (fields(3).toLong % 10, Rating(fields(0).toInt, fields(1).toInt, fields(2).toDouble))

    }

    // 装载电影目录对照表（电影ID->电影标题）

    val movies = sc.textFile(new File(movieLensHomeDir, "movies.dat").toString).map { line =>

      val fields = line.split("::")

      (fields(0).toInt, fields(1))

    }.collect().toMap

    val numRatings = ratings.count()

    val numUsers = ratings.map(_._2.user).distinct().count()

    val numMovies = ratings.map(_._2.product).distinct().count()

    println("Got " + numRatings + " ratings from " + numUsers + " users on " + numMovies + " movies.")

    // 将样本评分表以key值切分成3个部分，分别用于训练 (60%，并加入用户评分), 校验 (20%), and 测试 (20%)

    // 该数据在计算过程中要多次应用到，所以cache到内存

    val numPartitions = 4

    val training = ratings.filter(x => x._1 < 6)

      .values

      .union(myRatingsRDD) //注意ratings是(Int,Rating)，取value即可

      .repartition(numPartitions)

      .cache()

    val validation = ratings.filter(x => x._1 >= 6 && x._1 < 8)

      .values

      .repartition(numPartitions)

      .cache()

    val test = ratings.filter(x => x._1 >= 8).values.cache()

    val numTraining = training.count()

    val numValidation = validation.count()

    val numTest = test.count()

    println("Training: " + numTraining + ", validation: " + numValidation + ", test: " + numTest)

    // 训练不同参数下的模型，并在校验集中验证，获取最佳参数下的模型

    val ranks = List(8, 12)

    val lambdas = List(0.1, 10.0)

    val numIters = List(10, 20)

    var bestModel: Option[MatrixFactorizationModel] = None

    var bestValidationRmse = Double.MaxValue

    var bestRank = 0

    var bestLambda = -1.0

    var bestNumIter = -1

    for (rank <- ranks; lambda <- lambdas; numIter <- numIters) {

      val model = ALS.train(training, rank, numIter, lambda)

      val validationRmse = computeRmse(model, validation, numValidation)

      println("RMSE (validation) = " + validationRmse + " for the model trained with rank = "

        + rank + ", lambda = " + lambda + ", and numIter = " + numIter + ".")

      if (validationRmse < bestValidationRmse) {

        bestModel = Some(model)

        bestValidationRmse = validationRmse

        bestRank = rank

        bestLambda = lambda

        bestNumIter = numIter

      }

    }

    // 用最佳模型预测测试集的评分，并计算和实际评分之间的均方根误差

    val testRmse = computeRmse(bestModel.get, test, numTest)

    println("The best model was trained with rank = " + bestRank + " and lambda = " + bestLambda  + ", and numIter = " + bestNumIter + ", and its RMSE on the test set is " + testRmse + ".")

    // create a naive baseline and compare it with the best model

    val meanRating = training.union(validation).map(_.rating).mean

    val baselineRmse =

      math.sqrt(test.map(x => (meanRating - x.rating) * (meanRating - x.rating)).mean)

    val improvement = (baselineRmse - testRmse) / baselineRmse * 100

    println("The best model improves the baseline by " + "%1.2f".format(improvement) + "%.")

    // 推荐前十部最感兴趣的电影，注意要剔除用户已经评分的电影

    val myRatedMovieIds = myRatings.map(_.product).toSet

    val candidates = sc.parallelize(movies.keys.filter(!myRatedMovieIds.contains(_)).toSeq)

    val recommendations = bestModel.get

      .predict(candidates.map((0, _)))

      .collect()

      .sortBy(-_.rating)

      .take(10)

    var i = 1

    println("Movies recommended for you:")

    recommendations.foreach { r =>

      println("%2d".format(i) + ": " + movies(r.product))

      i += 1

    }

  sc.stop()

  }

  /** 校验集预测数据和实际数据之间的均方根误差 **/

  def computeRmse(model: MatrixFactorizationModel, data: RDD[Rating], n: Long): Double = {

    val predictions: RDD[Rating] = model.predict(data.map(x => (x.user, x.product)))

    val predictionsAndRatings = predictions.map(x => ((x.user, x.product), x.rating))

      .join(data.map(x => ((x.user, x.product), x.rating)))

      .values

    math.sqrt(predictionsAndRatings.map(x => (x._1 - x._2) * (x._1 - x._2)).reduce(_ + _) / n)

  }

  /** 装载用户评分文件 **/

  def loadRatings(path: String): Seq[Rating] = {

    val lines = Source.fromFile(path).getLines()

    val ratings = lines.map { line =>

      val fields = line.split("::")

      Rating(fields(0).toInt, fields(1).toInt, fields(2).toDouble)

    }.filter(_.rating > 0.0)

    if (ratings.isEmpty) {

      sys.error("No ratings provided.")

    } else {

      ratings.toSeq

    }

  }

}