回归

回归问题其实就是求解一堆自变量与因变量之间一种几何关系,这种关系可以是线性的就是线性回归,可以是非线性的就是非线性回归。按照自变量的多少有可以分为一元线性回归,多元线性回归。

线性回归

线性回归,顾名思义拟合出来的预测函数是一条直线,数学表达如下: h(x)=a0+a1x1+a2x2+…+anxn+J(θ) 其中 h(x)为预测函数, ai(i=1,2,…,n)为估计参数,模型训练的目的就是计算出这些参数的值。而线性回归分析的整个过程可以简单描述为如下三个步骤:

寻找合适的预测函数,即上文中的h(x)用来预测输入数据的判断结果。这个过程时非常关键的,需要对数据有一定的了解或分析,知道或者猜测预测函数的“大概”形式,比如是线性函数还是非线性函数,若是非线性的则无法用线性回归来得出高质量的结果。

构造一个Loss函数(损失函数),该函数表示预测的输出(h)与训练数据标签之间的偏差,可以是二者之间的差(h-y)或者是其他的形式(如平方差开方)。综合考虑所有训练数据的“损失”,将Loss求和或者求平均,记为J(θ)函数,表示所有训练数据预测值与实际类别的偏差。

显然, J(θ) 函数的值越小表示预测函数越准确(即h函数越准确),所以这一步需要做的是找到 J(θ)函数的最小值。找函数的最小值有不同的方法,Spark中采用的是随机梯度下降法(stochastic gradient descent SGD)。

线性回归同样可以采用正则化手段,其主要目的就是防止过拟合。MLlib提供L1、L2和无正则化三种方法: 当采用L1正则化时,则变成了Lasso Regresion;当采用L2正则化时,则变成了Ridge Regression;线性回归未采用正则化手段。通常来说,在训练模型时是建议采用正则化手段的,特别是在训练数据的量特别少的时候,若不采用正则化手段,过拟合现象会非常严重。L2正则化相比L1而言会更容易收敛(迭代次数少),但L1可以解决训练数据量小于维度的问题(也就是n元一次方程只有不到n个表达式,这种情况下是多解或无穷解的)。 例子:

import org.apache.spark.ml.regression.LinearRegression

// Load training data
val training = spark.read.format("libsvm")
  .load("data/mllib/sample_linear_regression_data.txt")

val lr = new LinearRegression()
  .setMaxIter(10)
  .setRegParam(0.3)
  .setElasticNetParam(0.8)

// Fit the model
val lrModel = lr.fit(training)

// Print the coefficients and intercept for linear regression
println(s"Coefficients: ${lrModel.coefficients} Intercept: ${lrModel.intercept}")

// Summarize the model over the training set and print out some metrics
val trainingSummary = lrModel.summary
println(s"numIterations: ${trainingSummary.totalIterations}")
println(s"objectiveHistory: [${trainingSummary.objectiveHistory.mkString(",")}]")
trainingSummary.residuals.show()
println(s"RMSE: ${trainingSummary.rootMeanSquaredError}")
println(s"r2: ${trainingSummary.r2}")
广义线性回归

与线性回归假设输出服从高斯分布不同,广义线性模型(GLMs)指定线性模型的因变量 服从指数型分布。Spark的GeneralizedLinearRegression接口允许指定GLMs包括线性回归、泊松回归、逻辑回归等来处理多种预测问题。目前 spark.ml仅支持指数型分布家族中的一部分类型,如下:

经典的深度学习回归预测模型 回归问题预测的算法_lua

例子:

import org.apache.spark.ml.regression.GeneralizedLinearRegression

// Load training data
val dataset = spark.read.format("libsvm")
  .load("data/mllib/sample_linear_regression_data.txt")

val glr = new GeneralizedLinearRegression()
  .setFamily("gaussian")
  .setLink("identity")
  .setMaxIter(10)
  .setRegParam(0.3)

// Fit the model
val model = glr.fit(dataset)

// Print the coefficients and intercept for generalized linear regression model
println(s"Coefficients: ${model.coefficients}")
println(s"Intercept: ${model.intercept}")

// Summarize the model over the training set and print out some metrics
val summary = model.summary
println(s"Coefficient Standard Errors: ${summary.coefficientStandardErrors.mkString(",")}")
println(s"T Values: ${summary.tValues.mkString(",")}")
println(s"P Values: ${summary.pValues.mkString(",")}")
println(s"Dispersion: ${summary.dispersion}")
println(s"Null Deviance: ${summary.nullDeviance}")
println(s"Residual Degree Of Freedom Null: ${summary.residualDegreeOfFreedomNull}")
println(s"Deviance: ${summary.deviance}")
println(s"Residual Degree Of Freedom: ${summary.residualDegreeOfFreedom}")
println(s"AIC: ${summary.aic}")
println("Deviance Residuals: ")
summary.residuals().show()
决策树回归

理论请看SparkML(三)的决策树分类器 例子:

import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.evaluation.RegressionEvaluator
import org.apache.spark.ml.feature.VectorIndexer
import org.apache.spark.ml.regression.DecisionTreeRegressionModel
import org.apache.spark.ml.regression.DecisionTreeRegressor

// Load the data stored in LIBSVM format as a DataFrame.
val data = spark.read.format("libsvm").load("data/mllib/sample_libsvm_data.txt")

// Automatically identify categorical features, and index them.
// Here, we treat features with > 4 distinct values as continuous.
val featureIndexer = new VectorIndexer()
  .setInputCol("features")
  .setOutputCol("indexedFeatures")
  .setMaxCategories(4)
  .fit(data)

// Split the data into training and test sets (30% held out for testing).
val Array(trainingData, testData) = data.randomSplit(Array(0.7, 0.3))

// Train a DecisionTree model.
val dt = new DecisionTreeRegressor()
  .setLabelCol("label")
  .setFeaturesCol("indexedFeatures")

// Chain indexer and tree in a Pipeline.
val pipeline = new Pipeline()
  .setStages(Array(featureIndexer, dt))

// Train model. This also runs the indexer.
val model = pipeline.fit(trainingData)

// Make predictions.
val predictions = model.transform(testData)

// Select example rows to display.
predictions.select("prediction", "label", "features").show(5)

// Select (prediction, true label) and compute test error.
val evaluator = new RegressionEvaluator()
  .setLabelCol("label")
  .setPredictionCol("prediction")
  .setMetricName("rmse")
val rmse = evaluator.evaluate(predictions)
println(s"Root Mean Squared Error (RMSE) on test data = $rmse")

val treeModel = model.stages(1).asInstanceOf[DecisionTreeRegressionModel]
println(s"Learned regression tree model:\n ${treeModel.toDebugString}")
随机森林回归

理论请看SparkML(三)的随机森林 例子:

import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.evaluation.RegressionEvaluator
import org.apache.spark.ml.feature.VectorIndexer
import org.apache.spark.ml.regression.{RandomForestRegressionModel, RandomForestRegressor}

// Load and parse the data file, converting it to a DataFrame.
val data = spark.read.format("libsvm").load("data/mllib/sample_libsvm_data.txt")

// Automatically identify categorical features, and index them.
// Set maxCategories so features with > 4 distinct values are treated as continuous.
val featureIndexer = new VectorIndexer()
  .setInputCol("features")
  .setOutputCol("indexedFeatures")
  .setMaxCategories(4)
  .fit(data)

// Split the data into training and test sets (30% held out for testing).
val Array(trainingData, testData) = data.randomSplit(Array(0.7, 0.3))

// Train a RandomForest model.
val rf = new RandomForestRegressor()
  .setLabelCol("label")
  .setFeaturesCol("indexedFeatures")

// Chain indexer and forest in a Pipeline.
val pipeline = new Pipeline()
  .setStages(Array(featureIndexer, rf))

// Train model. This also runs the indexer.
val model = pipeline.fit(trainingData)

// Make predictions.
val predictions = model.transform(testData)

// Select example rows to display.
predictions.select("prediction", "label", "features").show(5)

// Select (prediction, true label) and compute test error.
val evaluator = new RegressionEvaluator()
  .setLabelCol("label")
  .setPredictionCol("prediction")
  .setMetricName("rmse")
val rmse = evaluator.evaluate(predictions)
println(s"Root Mean Squared Error (RMSE) on test data = $rmse")

val rfModel = model.stages(1).asInstanceOf[RandomForestRegressionModel]
println(s"Learned regression forest model:\n ${rfModel.toDebugString}")
梯度提升树回归

理论请看SparkML(三)的梯度提升树分类器 例子:

import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.evaluation.RegressionEvaluator
import org.apache.spark.ml.feature.VectorIndexer
import org.apache.spark.ml.regression.{GBTRegressionModel, GBTRegressor}

// Load and parse the data file, converting it to a DataFrame.
val data = spark.read.format("libsvm").load("data/mllib/sample_libsvm_data.txt")

// Automatically identify categorical features, and index them.
// Set maxCategories so features with > 4 distinct values are treated as continuous.
val featureIndexer = new VectorIndexer()
  .setInputCol("features")
  .setOutputCol("indexedFeatures")
  .setMaxCategories(4)
  .fit(data)

// Split the data into training and test sets (30% held out for testing).
val Array(trainingData, testData) = data.randomSplit(Array(0.7, 0.3))

// Train a GBT model.
val gbt = new GBTRegressor()
  .setLabelCol("label")
  .setFeaturesCol("indexedFeatures")
  .setMaxIter(10)

// Chain indexer and GBT in a Pipeline.
val pipeline = new Pipeline()
  .setStages(Array(featureIndexer, gbt))

// Train model. This also runs the indexer.
val model = pipeline.fit(trainingData)

// Make predictions.
val predictions = model.transform(testData)

// Select example rows to display.
predictions.select("prediction", "label", "features").show(5)

// Select (prediction, true label) and compute test error.
val evaluator = new RegressionEvaluator()
  .setLabelCol("label")
  .setPredictionCol("prediction")
  .setMetricName("rmse")
val rmse = evaluator.evaluate(predictions)
println(s"Root Mean Squared Error (RMSE) on test data = $rmse")

val gbtModel = model.stages(1).asInstanceOf[GBTRegressionModel]
println(s"Learned regression GBT model:\n ${gbtModel.toDebugString}")
生存回归

在生存分析领域,加速失效时间模型(accelerated failure time model,AFT 模型)可以作为比例风险模型的替代模型。AFT模型将线性回归模型的建模方法引人到生存分析的领域, 将生存时间的对数作为反应变量,研究多协变量与对数生存时间之间的回归关系,在形式上,模型与一般的线性回归模型相似。对回归系数的解释也与一般的线性回归模型相似,较之Cox模型, AFT模型对分析结果的解释更加简单、直观且易于理解,并且可以预测个体的生存时间。 在中spark.ml,我们实现了加速故障时间(AFT) 模型,该模型是用于审查数据的参数生存回归模型。它描述了生存时间的对数模型,因此通常称为生存分析的对数线性模型。与为相同目的设计的比例风险模型不同 ,AFT模型更易于并行化,因为每个实例独立地对目标函数做出贡献。详情可以参考这篇博客 例子:

import org.apache.spark.ml.linalg.Vectors
import org.apache.spark.ml.regression.AFTSurvivalRegression

val training = spark.createDataFrame(Seq(
  (1.218, 1.0, Vectors.dense(1.560, -0.605)),
  (2.949, 0.0, Vectors.dense(0.346, 2.158)),
  (3.627, 0.0, Vectors.dense(1.380, 0.231)),
  (0.273, 1.0, Vectors.dense(0.520, 1.151)),
  (4.199, 0.0, Vectors.dense(0.795, -0.226))
)).toDF("label", "censor", "features")
val quantileProbabilities = Array(0.3, 0.6)
val aft = new AFTSurvivalRegression()
  .setQuantileProbabilities(quantileProbabilities)
  .setQuantilesCol("quantiles")

val model = aft.fit(training)

// Print the coefficients, intercept and scale parameter for AFT survival regression
println(s"Coefficients: ${model.coefficients}")
println(s"Intercept: ${model.intercept}")
println(s"Scale: ${model.scale}")
model.transform(training).show(false)
等渗回归

等渗回归 属于回归算法家族。形式等渗回归是一个问题,要给定一个有限的实数集,这些实数集代表观察到的响应,而未知的响应值需要拟合,从而找到一个最小化的函数 我们实现了一个 池违反者算法 ,该算法使用一种 等渗回归并行化方法。训练输入是一个DataFrame,其中包含三列标签,特征和权重。此外,IsoniconicRegression算法具有一个可选参数,称为默认为true。该参数指定等渗回归是等渗(单调增加)还是反渗(单调减少)。 训练返回一个IsotonicRegressionModel,该模型可用于预测已知和未知特征的标签。等渗回归的结果被视为分段线性函数。因此,预测规则为:

  • 如果预测输入与训练特征完全匹配,则返回关联的预测。如果有多个具有相同特征的预测,则返回其中之一。哪一个未定义(与java.util.Arrays.binarySearch相同)。
  • 如果预测输入低于或高于所有训练特征,则分别返回具有最低或最高特征的预测。如果有多个具有相同特征的预测,则分别返回最低或最高。
  • 如果预测输入介于两个训练特征之间,则将预测视为分段线性函数,并根据两个最接近特征的预测来计算内插值。如果有多个具有相同特征的值,则使用与上一点相同的规则。 例子:
import org.apache.spark.ml.regression.IsotonicRegression

// Loads data.
val dataset = spark.read.format("libsvm")
  .load("data/mllib/sample_isotonic_regression_libsvm_data.txt")

// Trains an isotonic regression model.
val ir = new IsotonicRegression()
val model = ir.fit(dataset)

println(s"Boundaries in increasing order: ${model.boundaries}\n")
println(s"Predictions associated with the boundaries: ${model.predictions}\n")

// Makes predictions.
model.transform(dataset).show()
分解机回归器

FM主要是解决稀疏数据下的特征组合问题,并且其预测的复杂度是线性的,对于连续和离散特征有较好的通用性。详细请参考 例子:

import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.evaluation.RegressionEvaluator
import org.apache.spark.ml.feature.MinMaxScaler
import org.apache.spark.ml.regression.{FMRegressionModel, FMRegressor}

// Load and parse the data file, converting it to a DataFrame.
val data = spark.read.format("libsvm").load("data/mllib/sample_libsvm_data.txt")

// Scale features.
val featureScaler = new MinMaxScaler()
  .setInputCol("features")
  .setOutputCol("scaledFeatures")
  .fit(data)

// Split the data into training and test sets (30% held out for testing).
val Array(trainingData, testData) = data.randomSplit(Array(0.7, 0.3))

// Train a FM model.
val fm = new FMRegressor()
  .setLabelCol("label")
  .setFeaturesCol("scaledFeatures")
  .setStepSize(0.001)

// Create a Pipeline.
val pipeline = new Pipeline()
  .setStages(Array(featureScaler, fm))

// Train model.
val model = pipeline.fit(trainingData)

// Make predictions.
val predictions = model.transform(testData)

// Select example rows to display.
predictions.select("prediction", "label", "features").show(5)

// Select (prediction, true label) and compute test error.
val evaluator = new RegressionEvaluator()
  .setLabelCol("label")
  .setPredictionCol("prediction")
  .setMetricName("rmse")
val rmse = evaluator.evaluate(predictions)
println(s"Root Mean Squared Error (RMSE) on test data = $rmse")

val fmModel = model.stages(1).asInstanceOf[FMRegressionModel]
println(s"Factors: ${fmModel.factors} Linear: ${fmModel.linear} " +
  s"Intercept: ${fmModel.intercept}")