SparkML之推荐算法（一）ALS
来源：互联网 发布：淘宝兼职红包单 编辑：程序博客网 时间：2024/04/23 20:14
ALS(alternating least squares ):交替最小二乘法
---------------------------------------------------------------------
原理应用
Matlab 主成分分析应用als
Spark源码
SparkML实验
---------------------------------------------------------------------
ALS在推荐系统中应用参考论文：见文献1
对分布式原理兴趣的还可以读读這篇文章：见文献2
原理
下面从文献1中取材，来讲解这个交替最小二乘法在推荐系统中应用的问题。如下图，对于一个R（观众对电影
的一个评价矩阵）可以分解为U（观众的特征矩阵）和V（电影的特征矩阵）
现在假如观众有5个人，电影有5部，那么R就是一个5*5的矩阵。假设评分如下：
假设d是三个属性（性格，文化程度，兴趣爱好）那么U的矩阵如下：
V的矩阵如下：
仔细的人或许发现是R约等于U*V，为什么是约等于呢？因为对于一个U矩阵来说，我们并不可能说（性格，文化程
度，兴趣爱好）这三个属性就代表着一个人对一部电影评价全部的属性，比如还有地域等因素。但是我们可以用“主
成分分析的思想”来近似（我没有从纯数学角度来谈，是为了大家更好理解）。這也是ALS和核心：一个评分矩阵可
以用两个小矩阵来近似（ALS是NNMF问题下在丢失数据情况下的一个重要手段）。
那么如何评价这两个矩阵的好坏？
理想的情况下：
但是实际中我们求使用数值计算的方法来求，那么计算得到的
就会存在误差，如何评价
的好坏。采用如下
表达
其中RMSE的意思是均方根误差（root mean square error），是u对v评分的预测值，
是u对v评分的观察值。
的表达如下：
那么就是转化的要求和
和
，
现在出现了一共有
个参数需要求解，而且碰到以下问题：
个参数需要求解，而且碰到以下问题：（1）、我们所知道的K矩阵是稀疏矩阵（K就是R在U对V没有全部评价的矩阵）
（2）、K的大小远小于远小于R的密集对应的大小
在解决这样的一个问题是采用拟合数据的形式来进行解决数据是稀疏的问题，公式如下：
note：其实后面的
是为了解决过拟合问题而增加的。
是为了解决过拟合问题而增加的。对于ALS来求解這样這个问题的思想是：先固定或者
,然后就转化为最小二乘法的问题了。他這样做就可以把一个非凸函数的问题转为二次函数的问题了。下面就求解步骤[1]：
步骤1：初始化矩阵V（可以取平均值也可以随机取值）
步骤1：初始化矩阵V（可以取平均值也可以随机取值）
步骤2：固定V，然后通过最小化误差函数(RMSE)解决求解U
步骤3：固定步骤2中的U，然后通过最小化误差函数(RMSE)解决求解V
步骤4：反复步骤2，3；直到U和V收敛。
梳理：为什么是交替，从处理步骤来看就是确定V，来优化U，再来优化V，再来优化U，。。。。直到收敛
因为采用梯度下降和最小二乘都可以解决這个问题，在此不写代码来讲如何决定参数，可以看前面的最小二乘或者梯度下降算法。下面结合matlab来展示一下，利用主成分分析，借助ALS来完善丢失数据来预测电影评分的效果：
数据：链接：http://pan.baidu.com/s/1kUWo2iF 密码：iie8
%目的% 通过主成分分析，用ALS来优化，同时来得到潜在的评分，数据就是上面观众看电影数据load Data.txtR = Data;[coeff1,score1,latent,tsquared,explained,mu1] = pca(R,'algorithm','als');%% 参数%coeff1 主成分系数% 0.2851 -0.5043 0.8124 -0.0266% 0.9230 -0.0764 -0.3655 0.0830% -0.1822 -0.4338 -0.1826 0.8602% -0.0890 -0.2844 -0.0782 -0.0986% 0.1602 0.6861 0.4085 0.4927%score1 主成分得分% 3.1434 -2.0913 -0.1917 -0.0505% -3.1122 0.5615 -0.1839 -0.2997% -4.9612 -0.4934 -0.0388 0.2334% 3.3091 1.5365 -0.4941 0.1154% 1.6210 0.4868 0.9086 0.0014%latent 主成分方差%14.4394% 1.8826% 0.2854% 0.0400%tsquared Hotelling的T平方统计，在X每个观测%3.2000% 3.2000% 3.2000% 3.2000% 3.2000%explained 向量包含每个主成分解释的总方差的百分比% 86.7366% 11.3084% 1.7145% 0.2405%mu1 返回的平均值% 5.2035 3.8730 4.6740 4.7043 5.0344%% 重建矩阵（预测）p = score1*coeff1' + repmat(mu1,5,1)%7.0000 7.0000 5.0000 5.0393 4.0000%3.8915 1.0000 4.7733 4.8655 4.6982%4.0000 -0.6348 6.0000 5.2662 4.0000% 4.9677 7.0000 3.5939 4.0000 6.4738%6.1583 5.0000 4.0027 4.3503 6.0000
再来分析一下Spark对ALS优化参数部分，这部分因为“观看者”和“电影”数据进行了block化（源码对其两个参数命名为：User和products，为了和源码一样，所以接下来的分析用User和products这两个名字），所以拉来重点说：
官方参考文献参考的是文献3，它的核心思想是,对User数据和products数据进行Block化，目的：减少数据通信。
为了让大家清楚block可以带减少数据通信的优势，我极端化分析一下，也就是把上面每个用户和每个电影都block
如下：
从上图可以发现，在进行参数优化确定V，来优化U，再来优化V，再来优化U，。。。。直到收敛，这样的一
个过程，对数据进行Block化确实可以减少通信消耗。
源码
/** * 一个比Tuple3[Int, Int, Double]更加紧凑的class 用于表示评分 */@Since("0.8.0")case class Rating @Since("0.8.0") ( @Since("0.8.0") user: Int, @Since("0.8.0") product: Int, @Since("0.8.0") rating: Double)/** * 交替最小二乘法的矩阵分解。 * * ALS attempts to estimate the ratings matrix `R` as the product of two lower-rank matrices, * `X` and `Y`, i.e. `X * Yt = R`. Typically these approximations are called 'factor' matrices. * The general approach is iterative. During each iteration, one of the factor matrices is held * constant, while the other is solved for using least squares. The newly-solved factor matrix is * then held constant while solving for the other factor matrix. * * This is a blocked implementation of the ALS factorization algorithm that groups the two sets * of factors (referred to as "users" and "products") into blocks and reduces communication by only * sending one copy of each user vector to each product block on each iteration, and only for the * product blocks that need that user's feature vector. This is achieved by precomputing some * information about the ratings matrix to determine the "out-links" of each user (which blocks of * products it will contribute to) and "in-link" information for each product (which of the feature * vectors it receives from each user block it will depend on). This allows us to send only an * array of feature vectors between each user block and product block, and have the product block * find the users' ratings and update the products based on these messages. * * For implicit preference data, the algorithm used is based on * "Collaborative Filtering for Implicit Feedback Datasets", available at * [[http://dx.doi.org/10.1109/ICDM.2008.22]], adapted for the blocked approach used here. * * Essentially instead of finding the low-rank approximations to the rating matrix `R`, * this finds the approximations for a preference matrix `P` where the elements of `P` are 1 if * r > 0 and 0 if r <= 0. The ratings then act as 'confidence' values related to strength of * indicated user * preferences rather than explicit ratings given to items. */@Since("0.8.0")class ALS private ( private var numUserBlocks: Int, private var numProductBlocks: Int, private var rank: Int, private var iterations: Int, private var lambda: Double, private var implicitPrefs: Boolean, private var alpha: Double, private var seed: Long = System.nanoTime() ) extends Serializable with Logging { /** * 构造一个默认参数的ALS的实例: {numBlocks: -1, rank: 10, iterations: 10, * lambda: 0.01, implicitPrefs: false, alpha: 1.0}. */ @Since("0.8.0") def this() = this(-1, -1, 10, 10, 0.01, false, 1.0) /** 如果是true，做交替的非负最小二乘。. */ private var nonnegative = false /** storage level for user/product in/out links */ private var intermediateRDDStorageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK private var finalRDDStorageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK /** checkpoint interval */ private var checkpointInterval: Int = 10 /** * 设置用户模块和产品模块并行计算块的数量（假设设置为2，那么用户和产品的模块都是2个）,numBlocks=-1的时候表示自动配置模块数，默认情况下是 numBlocks=-1 */ @Since("0.8.0") def setBlocks(numBlocks: Int): this.type = { require(numBlocks == -1 || numBlocks > 0, s"Number of blocks must be -1 or positive but got ${numBlocks}") this.numUserBlocks = numBlocks this.numProductBlocks = numBlocks this } /** * 设置并行计算的用户的块的数量。 */ @Since("1.1.0") def setUserBlocks(numUserBlocks: Int): this.type = { require(numUserBlocks == -1 || numUserBlocks > 0, s"Number of blocks must be -1 or positive but got ${numUserBlocks}") this.numUserBlocks = numUserBlocks this } /** * 设置并行计算的产品的块的数量。 */ @Since("1.1.0") def setProductBlocks(numProductBlocks: Int): this.type = { require(numProductBlocks == -1 || numProductBlocks > 0, s"Number of product blocks must be -1 or positive but got ${numProductBlocks}") this.numProductBlocks = numProductBlocks this } /** 计算特征矩阵的秩（特征数），默认情况下为10 */ @Since("0.8.0") def setRank(rank: Int): this.type = { require(rank > 0, s"Rank of the feature matrices must be positive but got ${rank}") this.rank = rank this } /** 设置要运行的迭代次数。默人为10次 */ @Since("0.8.0") def setIterations(iterations: Int): this.type = { require(iterations >= 0, s"Number of iterations must be nonnegative but got ${iterations}") this.iterations = iterations this } /** 集的正则化参数，λ. 默认为 0.01. */ @Since("0.8.0") def setLambda(lambda: Double): this.type = { require(lambda >= 0.0, s"Regularization parameter must be nonnegative but got ${lambda}") this.lambda = lambda this } /** 设置是否使用隐式偏好。Default: false. */ @Since("0.8.1") def setImplicitPrefs(implicitPrefs: Boolean): this.type = { this.implicitPrefs = implicitPrefs this } /** * Sets the constant used in computing confidence in implicit ALS. Default: 1.0. */ @Since("0.8.1") def setAlpha(alpha: Double): this.type = { this.alpha = alpha this } /** Sets a random seed to have deterministic results. */ @Since("1.0.0") def setSeed(seed: Long): this.type = { this.seed = seed this } /** * * * 设置每一次迭代中的最小二乘法，是否都要非负约束 * Set whether the least-squares problems solved at each iteration should have * nonnegativity constraints. */ @Since("1.1.0") def setNonnegative(b: Boolean): this.type = { this.nonnegative = b this } /** * :: DeveloperApi ::* 对每一个RDD在中间的缓存级别的选择 * Sets storage level for intermediate RDDs (user/product in/out links). The default value is * `MEMORY_AND_DISK`. Users can change it to a serialized storage, e.g., `MEMORY_AND_DISK_SER` and * set `spark.rdd.compress` to `true` to reduce the space requirement, at the cost of speed. */ @DeveloperApi @Since("1.1.0") def setIntermediateRDDStorageLevel(storageLevel: StorageLevel): this.type = { require(storageLevel != StorageLevel.NONE, "ALS is not designed to run without persisting intermediate RDDs.") this.intermediateRDDStorageLevel = storageLevel this } /** * :: DeveloperApi :: * Sets storage level for final RDDs (user/product used in MatrixFactorizationModel). The default * value is `MEMORY_AND_DISK`. Users can change it to a serialized storage, e.g. * `MEMORY_AND_DISK_SER` and set `spark.rdd.compress` to `true` to reduce the space requirement, * at the cost of speed. */ @DeveloperApi @Since("1.3.0") def setFinalRDDStorageLevel(storageLevel: StorageLevel): this.type = { this.finalRDDStorageLevel = storageLevel this } /** * Set period (in iterations) between checkpoints (default = 10). Checkpointing helps with * recovery (when nodes fail) and StackOverflow exceptions caused by long lineage. It also helps * with eliminating temporary shuffle files on disk, which can be important when there are many * ALS iterations. If the checkpoint directory is not set in [[org.apache.spark.SparkContext]], * this setting is ignored. */ @DeveloperApi @Since("1.4.0") //设置每隔多久 checkpoint一下 def setCheckpointInterval(checkpointInterval: Int): this.type = { this.checkpointInterval = checkpointInterval this } /** * Run ALS with the configured parameters on an input RDD of [[Rating]] objects. * Returns a MatrixFactorizationModel with feature vectors for each user and product. */ @Since("0.8.0") //查看一开始给的Rating,這个RDD的形式内部数据如下： //case class Rating @Since("0.8.0") ( // @Since("0.8.0") user: Int, // @Since("0.8.0") product: Int, // @Since("0.8.0") rating: Double) def run(ratings: RDD[Rating]): MatrixFactorizationModel = { val sc = ratings.context //分块设置，默认下：在并行度和rating的partitions的二分之一中选一个最大的 // 设置参数下：为numUserBlocks val numUserBlocks = if (this.numUserBlocks == -1) { math.max(sc.defaultParallelism, ratings.partitions.length / 2) } else { this.numUserBlocks } //分块设置，默认下：在并行度和rating的partitions的二分之一中选一个最大的 // 设置参数下：为numProductBlocks val numProductBlocks = if (this.numProductBlocks == -1) { math.max(sc.defaultParallelism, ratings.partitions.length / 2) } else { this.numProductBlocks } val (floatUserFactors, floatProdFactors) = NewALS.train[Int]( ratings = ratings.map(r => NewALS.Rating(r.user, r.product, r.rating.toFloat)), rank = rank, numUserBlocks = numUserBlocks, numItemBlocks = numProductBlocks, maxIter = iterations, regParam = lambda, implicitPrefs = implicitPrefs, alpha = alpha, nonnegative = nonnegative, intermediateRDDStorageLevel = intermediateRDDStorageLevel, finalRDDStorageLevel = StorageLevel.NONE, checkpointInterval = checkpointInterval, seed = seed) val userFactors = floatUserFactors .mapValues(_.map(_.toDouble)) .setName("users") .persist(finalRDDStorageLevel) val prodFactors = floatProdFactors .mapValues(_.map(_.toDouble)) .setName("products") .persist(finalRDDStorageLevel) if (finalRDDStorageLevel != StorageLevel.NONE) { userFactors.count() prodFactors.count() } new MatrixFactorizationModel(rank, userFactors, prodFactors) } /** * Java-friendly version of [[ALS.run]]. */ @Since("1.3.0") def run(ratings: JavaRDD[Rating]): MatrixFactorizationModel = run(ratings.rdd)}/** * Top-level methods for calling Alternating Least Squares (ALS) matrix factorization. */@Since("0.8.0")object ALS { /** * Train a matrix factorization model given an RDD of ratings by users for a subset of products. * The ratings matrix is approximated as the product of two lower-rank matrices of a given rank * (number of features). To solve for these features, ALS is run iteratively with a configurable * level of parallelism. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter * @param blocks level of parallelism to split computation into * @param seed random seed for initial matrix factorization model */ @Since("0.9.1") def train( ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double, blocks: Int, seed: Long ): MatrixFactorizationModel = { new ALS(blocks, blocks, rank, iterations, lambda, false, 1.0, seed).run(ratings) } /** * Train a matrix factorization model given an RDD of ratings by users for a subset of products. * The ratings matrix is approximated as the product of two lower-rank matrices of a given rank * (number of features). To solve for these features, ALS is run iteratively with a configurable * level of parallelism. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter * @param blocks level of parallelism to split computation into */ @Since("0.8.0") def train( ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double, blocks: Int ): MatrixFactorizationModel = { new ALS(blocks, blocks, rank, iterations, lambda, false, 1.0).run(ratings) } /** * Train a matrix factorization model given an RDD of ratings by users for a subset of products. * The ratings matrix is approximated as the product of two lower-rank matrices of a given rank * (number of features). To solve for these features, ALS is run iteratively with a level of * parallelism automatically based on the number of partitions in `ratings`. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter */ @Since("0.8.0") def train(ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double) : MatrixFactorizationModel = { train(ratings, rank, iterations, lambda, -1) } /** * Train a matrix factorization model given an RDD of ratings by users for a subset of products. * The ratings matrix is approximated as the product of two lower-rank matrices of a given rank * (number of features). To solve for these features, ALS is run iteratively with a level of * parallelism automatically based on the number of partitions in `ratings`. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS */ @Since("0.8.0") def train(ratings: RDD[Rating], rank: Int, iterations: Int) : MatrixFactorizationModel = { train(ratings, rank, iterations, 0.01, -1) } /** * Train a matrix factorization model given an RDD of 'implicit preferences' given by users * to some products, in the form of (userID, productID, preference) pairs. We approximate the * ratings matrix as the product of two lower-rank matrices of a given rank (number of features). * To solve for these features, we run a given number of iterations of ALS. This is done using * a level of parallelism given by `blocks`. * * @param ratings RDD of (userID, productID, rating) pairs * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter * @param blocks level of parallelism to split computation into * @param alpha confidence parameter * @param seed random seed for initial matrix factorization model */ @Since("0.8.1") def trainImplicit( ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double, blocks: Int, alpha: Double, seed: Long ): MatrixFactorizationModel = { new ALS(blocks, blocks, rank, iterations, lambda, true, alpha, seed).run(ratings) } /** * Train a matrix factorization model given an RDD of 'implicit preferences' of users for a * subset of products. The ratings matrix is approximated as the product of two lower-rank * matrices of a given rank (number of features). To solve for these features, ALS is run * iteratively with a configurable level of parallelism. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter * @param blocks level of parallelism to split computation into * @param alpha confidence parameter */ @Since("0.8.1") def trainImplicit( ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double, blocks: Int, alpha: Double ): MatrixFactorizationModel = { new ALS(blocks, blocks, rank, iterations, lambda, true, alpha).run(ratings) } /** * Train a matrix factorization model given an RDD of 'implicit preferences' of users for a * subset of products. The ratings matrix is approximated as the product of two lower-rank * matrices of a given rank (number of features). To solve for these features, ALS is run * iteratively with a level of parallelism determined automatically based on the number of * partitions in `ratings`. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS * @param lambda regularization parameter * @param alpha confidence parameter */ @Since("0.8.1") def trainImplicit(ratings: RDD[Rating], rank: Int, iterations: Int, lambda: Double, alpha: Double) : MatrixFactorizationModel = { trainImplicit(ratings, rank, iterations, lambda, -1, alpha) } /** * Train a matrix factorization model given an RDD of 'implicit preferences' of users for a * subset of products. The ratings matrix is approximated as the product of two lower-rank * matrices of a given rank (number of features). To solve for these features, ALS is run * iteratively with a level of parallelism determined automatically based on the number of * partitions in `ratings`. * * @param ratings RDD of [[Rating]] objects with userID, productID, and rating * @param rank number of features to use * @param iterations number of iterations of ALS */ @Since("0.8.1") def trainImplicit(ratings: RDD[Rating], rank: Int, iterations: Int) : MatrixFactorizationModel = { trainImplicit(ratings, rank, iterations, 0.01, -1, 1.0) }}MatrixFactorizationModel类
/** *矩阵分解模型 * * Note:如果你直接用构造函数来创建模型，请注意，快速预测需要缓存user/product, * * @param rank 秩 * @param userFeatures RDD的元组，每个元组都有计算后的userId 和 它features * @param productFeatures RDD的元组，每个元组都有计算后的productId 和 它features */@Since("0.8.0")class MatrixFactorizationModel @Since("0.8.0") ( @Since("0.8.0") val rank: Int, @Since("0.8.0") val userFeatures: RDD[(Int, Array[Double])], @Since("0.8.0") val productFeatures: RDD[(Int, Array[Double])]) extends Saveable with Serializable with Logging { require(rank > 0) validateFeatures("User", userFeatures) validateFeatures("Product", productFeatures) /**验证因素，并提醒用户，如果有性能问题。 */ private def validateFeatures(name: String, features: RDD[(Int, Array[Double])]): Unit = { require(features.first()._2.length == rank, s"$name feature dimension does not match the rank $rank.") if (features.partitioner.isEmpty) { logWarning(s"$name factor does not have a partitioner. " + "Prediction on individual records could be slow.") } if (features.getStorageLevel == StorageLevel.NONE) { logWarning(s"$name factor is not cached. Prediction could be slow.") } } /** 预测一个用户对一个产品的评价. */ @Since("0.8.0") def predict(user: Int, product: Int): Double = { val userVector = userFeatures.lookup(user).head val productVector = productFeatures.lookup(product).head blas.ddot(rank, userVector, 1, productVector, 1) } /** * 输入usersProducts，返回用户和产品的近似数，這个方法是基于countApproxDistinct * * @param usersProducts RDD of (user, product) pairs. * @return 用户和产品的近似值。 */ private[this] def countApproxDistinctUserProduct(usersProducts: RDD[(Int, Int)]): (Long, Long) = { val zeroCounterUser = new HyperLogLogPlus(4, 0) val zeroCounterProduct = new HyperLogLogPlus(4, 0) val aggregated = usersProducts.aggregate((zeroCounterUser, zeroCounterProduct))( (hllTuple: (HyperLogLogPlus, HyperLogLogPlus), v: (Int, Int)) => { hllTuple._1.offer(v._1) hllTuple._2.offer(v._2) hllTuple }, (h1: (HyperLogLogPlus, HyperLogLogPlus), h2: (HyperLogLogPlus, HyperLogLogPlus)) => { h1._1.addAll(h2._1) h1._2.addAll(h2._2) h1 }) (aggregated._1.cardinality(), aggregated._2.cardinality()) } /** * 预测多个用户对产品的评价。 *输出的RDD和输入的RDD元素一一对应 （包括所有副本）除非用户或产品中缺少训练集。 * @param usersProducts RDD of (user, product) pairs. * @return RDD of Ratings. */ @Since("0.9.0") def predict(usersProducts: RDD[(Int, Int)]): RDD[Rating] = { // Previously the partitions of ratings are only based on the given products. // So if the usersProducts given for prediction contains only few products or // even one product, the generated ratings will be pushed into few or single partition // and can't use high parallelism. // Here we calculate approximate numbers of users and products. Then we decide the // partitions should be based on users or products. val (usersCount, productsCount) = countApproxDistinctUserProduct(usersProducts) if (usersCount < productsCount) { val users = userFeatures.join(usersProducts).map { case (user, (uFeatures, product)) => (product, (user, uFeatures)) } users.join(productFeatures).map { case (product, ((user, uFeatures), pFeatures)) => Rating(user, product, blas.ddot(uFeatures.length, uFeatures, 1, pFeatures, 1)) } } else { val products = productFeatures.join(usersProducts.map(_.swap)).map { case (product, (pFeatures, user)) => (user, (product, pFeatures)) } products.join(userFeatures).map { case (user, ((product, pFeatures), uFeatures)) => Rating(user, product, blas.ddot(uFeatures.length, uFeatures, 1, pFeatures, 1)) } } } /** * Java-friendly version of [[MatrixFactorizationModel.predict]]. */ @Since("1.2.0") def predict(usersProducts: JavaPairRDD[JavaInteger, JavaInteger]): JavaRDD[Rating] = { predict(usersProducts.rdd.asInstanceOf[RDD[(Int, Int)]]).toJavaRDD() } /** * 向用户推荐产品。 * * @param user the user to recommend products to * @param num how many products to return. The number returned may be less than this. * @return [[Rating]] objects, each of which contains the given user ID, a product ID, and a * "score" in the rating field. Each represents one recommended product, and they are sorted * by score, decreasing. The first returned is the one predicted to be most strongly * recommended to the user. The score is an opaque value that indicates how strongly * recommended the product is. */ @Since("1.1.0") def recommendProducts(user: Int, num: Int): Array[Rating] = MatrixFactorizationModel.recommend(userFeatures.lookup(user).head, productFeatures, num) .map(t => Rating(user, t._1, t._2)) /** * 推荐用户给产品. 也就是说，看看那些用户对這个产品感兴趣 * * @param product 产品推荐用户 * @param num 设定返回多少个用户，实际返回的大小有可能少于设定的值 * @return [[Rating]] objects, 其中每一个包含用户的ID、产品ID，和一个得分，每个表示一个推荐的用户，并且它们按从大到小的分数排序，第一次返回的是一个预测是最建议的产品 */ @Since("1.1.0") def recommendUsers(product: Int, num: Int): Array[Rating] = MatrixFactorizationModel.recommend(productFeatures.lookup(product).head, userFeatures, num) .map(t => Rating(t._1, product, t._2)) protected override val formatVersion: String = "1.0" /** * 输入路径、保持模型 * * This saves: * - human-readable (JSON) model metadata to path/metadata/ * - Parquet formatted data to path/data/ * * The model may be loaded using [[Loader.load]]. * * @param sc Spark context used to save model data. * @param path Path specifying the directory in which to save this model. * If the directory already exists, this method throws an exception. */ @Since("1.3.0") override def save(sc: SparkContext, path: String): Unit = { MatrixFactorizationModel.SaveLoadV1_0.save(this, path) } /** * 为所有用户推荐top products。 * * @param num 为每个用户返回多少产品 * @return [(Int, Array[Rating])] objects, where every tuple contains a userID and an array of * rating objects which contains the same userId, recommended productID and a "score" in the * rating field. Semantics of score is same as recommendProducts API * */ @Since("1.4.0") def recommendProductsForUsers(num: Int): RDD[(Int, Array[Rating])] = { MatrixFactorizationModel.recommendForAll(rank, userFeatures, productFeatures, num).map { case (user, top) => val ratings = top.map { case (product, rating) => Rating(user, product, rating) } (user, ratings) } } /** * 为所有产品推荐 top users * * @param num how many users to return for every product. * @return [(Int, Array[Rating])] objects, where every tuple contains a productID and an array * of rating objects which contains the recommended userId, same productID and a "score" in the * rating field. Semantics of score is same as recommendUsers API */ @Since("1.4.0") def recommendUsersForProducts(num: Int): RDD[(Int, Array[Rating])] = { MatrixFactorizationModel.recommendForAll(rank, productFeatures, userFeatures, num).map { case (product, top) => val ratings = top.map { case (user, rating) => Rating(user, product, rating) } (product, ratings) } }}@Since("1.3.0")object MatrixFactorizationModel extends Loader[MatrixFactorizationModel] { import org.apache.spark.mllib.util.Loader._ /** * 对单个用户（或产品）进行推荐 */ private def recommend( recommendToFeatures: Array[Double], recommendableFeatures: RDD[(Int, Array[Double])], num: Int): Array[(Int, Double)] = { val scored = recommendableFeatures.map { case (id, features) => (id, blas.ddot(features.length, recommendToFeatures, 1, features, 1)) } scored.top(num)(Ordering.by(_._2)) } /** * 对所有用户（或产品）进行推荐 * @param rank rank * @param srcFeatures src features to receive recommendations * @param dstFeatures dst features used to make recommendations * @param num number of recommendations for each record * @return an RDD of (srcId: Int, recommendations), where recommendations are stored as an array * of (dstId, rating) pairs. */ private def recommendForAll( rank: Int, srcFeatures: RDD[(Int, Array[Double])], dstFeatures: RDD[(Int, Array[Double])], num: Int): RDD[(Int, Array[(Int, Double)])] = { val srcBlocks = blockify(rank, srcFeatures) val dstBlocks = blockify(rank, dstFeatures) val ratings = srcBlocks.cartesian(dstBlocks).flatMap { case ((srcIds, srcFactors), (dstIds, dstFactors)) => val m = srcIds.length val n = dstIds.length val ratings = srcFactors.transpose.multiply(dstFactors) val output = new Array[(Int, (Int, Double))](m * n) var k = 0 ratings.foreachActive { (i, j, r) => output(k) = (srcIds(i), (dstIds(j), r)) k += 1 } output.toSeq } ratings.topByKey(num)(Ordering.by(_._2)) } /** * Blockifies features to use Level-3 BLAS. */ private def blockify( rank: Int, features: RDD[(Int, Array[Double])]): RDD[(Array[Int], DenseMatrix)] = { val blockSize = 4096 // TODO: tune the block size val blockStorage = rank * blockSize features.mapPartitions { iter => iter.grouped(blockSize).map { grouped => val ids = mutable.ArrayBuilder.make[Int] ids.sizeHint(blockSize) val factors = mutable.ArrayBuilder.make[Double] factors.sizeHint(blockStorage) var i = 0 grouped.foreach { case (id, factor) => ids += id factors ++= factor i += 1 } (ids.result(), new DenseMatrix(rank, i, factors.result())) } } } /** * 输入模型的路径，加载這个模型 * * The model should have been saved by [[Saveable.save]]. * * @param sc Spark context used for loading model files. * @param path Path specifying the directory to which the model was saved. * @return Model instance */ @Since("1.3.0") override def load(sc: SparkContext, path: String): MatrixFactorizationModel = { val (loadedClassName, formatVersion, _) = loadMetadata(sc, path) val classNameV1_0 = SaveLoadV1_0.thisClassName (loadedClassName, formatVersion) match { case (className, "1.0") if className == classNameV1_0 => SaveLoadV1_0.load(sc, path) case _ => throw new IOException("MatrixFactorizationModel.load did not recognize model with" + s"(class: $loadedClassName, version: $formatVersion). Supported:\n" + s" ($classNameV1_0, 1.0)") } } private[recommendation] object SaveLoadV1_0 { private val thisFormatVersion = "1.0" private[recommendation] val thisClassName = "org.apache.spark.mllib.recommendation.MatrixFactorizationModel" /** * Saves a [[MatrixFactorizationModel]], where user features are saved under `data/users` and * product features are saved under `data/products`. */ def save(model: MatrixFactorizationModel, path: String): Unit = { val sc = model.userFeatures.sparkContext val sqlContext = SQLContext.getOrCreate(sc) import sqlContext.implicits._ val metadata = compact(render( ("class" -> thisClassName) ~ ("version" -> thisFormatVersion) ~ ("rank" -> model.rank))) sc.parallelize(Seq(metadata), 1).saveAsTextFile(metadataPath(path)) model.userFeatures.toDF("id", "features").write.parquet(userPath(path)) model.productFeatures.toDF("id", "features").write.parquet(productPath(path)) } def load(sc: SparkContext, path: String): MatrixFactorizationModel = { implicit val formats = DefaultFormats val sqlContext = SQLContext.getOrCreate(sc) val (className, formatVersion, metadata) = loadMetadata(sc, path) assert(className == thisClassName) assert(formatVersion == thisFormatVersion) val rank = (metadata \ "rank").extract[Int] val userFeatures = sqlContext.read.parquet(userPath(path)).rdd.map { case Row(id: Int, features: Seq[_]) => (id, features.asInstanceOf[Seq[Double]].toArray) } val productFeatures = sqlContext.read.parquet(productPath(path)).rdd.map { case Row(id: Int, features: Seq[_]) => (id, features.asInstanceOf[Seq[Double]].toArray) } new MatrixFactorizationModel(rank, userFeatures, productFeatures) } private def userPath(path: String): String = { new Path(dataPath(path), "user").toUri.toString } private def productPath(path: String): String = { new Path(dataPath(path), "product").toUri.toString } }}SparkML实验
import org.apache.log4j.{Level, Logger}import org.apache.spark.mllib.recommendation.{ALS, Rating}import org.apache.spark.{SparkConf, SparkContext}object myAls { def main(args: Array[String]) { val conf = new SparkConf().setAppName("Als example").setMaster("local[2]") val sc = new SparkContext(conf) Logger.getLogger("org.apache.spark").setLevel(Level.ERROR) Logger.getLogger("org.eclipse.jetty.Server").setLevel(Level.OFF) val trainData = sc.textFile("/root/application/upload/train.data") val parseTrainData =trainData.map(_.split(',') match{ case Array(user,item,rate) => Rating(user.toInt,item.toInt,rate.toDouble) }) val testData = sc.textFile("/root/application/upload/test.data") val parseTestData =testData.map(_.split(',') match{ case Array(user,item,rate) => Rating(user.toInt,item.toInt,rate.toDouble) }) parseTrainData.foreach(println) val model = new ALS().setBlocks(2).run(ratings = parseTrainData) val userProducts =parseTestData.map{ case Rating(user,product,rate) => (user,product) } val predictions = model.predict(userProducts).map{ case Rating(user,product,rate) => ((user,product),rate) } predictions.foreach(println) /** ((4,1),1.7896680396660953)((4,3),1.0270402568376826)((4,5),0.1556322625035942)((2,4),0.33505846168235803)((2,1),0.5416217248274381)((2,3),0.4346857699980956)((2,5),0.4549716283423277)((1,4),1.2289770624608378)((3,4),1.8560000519252107E-5)((3,2),3.3417571983500647)((5,4),-0.049730215285125445)((5,1),3.9938137663334397)((5,3),4.041703646645967) */ //预测的不怎么样 sc.stop() }}参考文献
http://cs229.stanford.edu/proj2014/Christopher%20Aberger,%20Recommender.pdf
http://www.grappa.univ-lille3.fr/~mary/cours/stats/centrale/reco/paper/MatrixFactorizationALS.pdf
http://yifanhu.net/PUB/cf.pdf
0 0
- SparkML之推荐算法（一）ALS
- SparkML之假设性检验（一）
- SparkML之DistributedMatrix(一)
- 基于矩阵分解推荐算法之交替最小二乘法（ALS）--附实现代码
- mahout之旅---分布式推荐算法ALS-MR
- Parallel-ALS推荐算法(factorize-movielens-1M)（转载）
- SparkML之特征提取（一）主成分分析(PCA)
- mahout0.9 分布式推荐算法ALS-MR
- Spark ALS算法推荐结果一样？
- Spark ML -1：ALS推荐算法
- 【推荐】基于Spark的ALS算法
- SparkML之回归(一)线性回归
- SparkML之分类(一)贝叶斯分类
- SparkML之聚类(一)Kmeans聚类
- 协同过滤之ALS-WR算法
- 协同过滤之ALS-WR算法
- 协同过滤之ALS-WR算法
- 基于ALS算法的简易在线推荐系统
- SparkML之聚类(二)高斯混合模型（GMMs）
- SparkML之特征提取（一）主成分分析(PCA)
- Swift编程规范之 Naming
- SparkML之特征提取（二）词项加权之DF-IDF
- 3497. 水仙花数
- SparkML之推荐算法（一）ALS
- 3 Sum
- IntelliJ IDEA快捷键大全（图片）
- NoSql入门和概述
- Random Number Series Questions
- SparkStreaming之基本数据源输入
- C++学习指南
- SparkStreaming之DStream operations
- SparkStreaming之foreachRDD
