标签:compute 分析 lan apach .class style vector stand nsf
SparkMLlib分类算法之决策树学习
(一) 决策树的基本概念
决策树(Decision Tree)是在已知各种情况发生概率的基础上,通过构成决策树来求取净现值的期望值大于等于零的概率,评价项目风险,判断其可行性的决策分析方法,是直观运用概率分析的一种图解法。由于这种决策分支画成图形很像一棵树的枝干,故称决策树。在机器学习中,决策树是一个预测模型,他代表的是对象属性与对象值之间的一种映射关系。Entropy = 系统的凌乱程度,使用算法ID3, C4.5和C5.0生成树算法使用熵。这一度量是基于信息学理论中熵的概念。通过信息增益来筛选出属性的优先性。
缺点:参考网址:http://www.ppvke.com/Blog/archives/25042
val orig_file=sc.textFile("train_nohead.tsv") //println(orig_file.first()) val data_file=orig_file.map(_.split("\t")).map{ r => val trimmed =r.map(_.replace("\"","")) val lable=trimmed(r.length-1).toDouble val feature=trimmed.slice(4,r.length-1).map(d => if(d=="?")0.0 else d.toDouble) LabeledPoint(lable,Vectors.dense(feature)) } /*特征标准化优化,似乎对决策数没啥影响*/ val vectors=data_file.map(x =>x.features) val rows=new RowMatrix(vectors) println(rows.computeColumnSummaryStatistics().variance)//每列的方差 val scaler=new StandardScaler(withMean=true,withStd=true).fit(vectors)//标准化 val scaled_data=data_file.map(point => LabeledPoint(point.label,scaler.transform(point.features))) .randomSplit(Array(0.7,0.3),11L)//固定seed为11L,确保每次每次实验能得到相同的结果 val data_train=scaled_data(0) val data_test=scaled_data(1)
3,构建模型及模型评价
/*训练决策树模型*/ val model_DT=DecisionTree.train(data_train,Algo.Classification,Entropy,maxTreeDepth) /*决策树的精确度*/ val predectionAndLabeledDT=data_test.map { point => val predectLabeled = if (model_DT.predict(point.features) > 0.5) 1.0 else 0.0 (predectLabeled,point.label) } val metricsDT=new MulticlassMetrics(predectionAndLabeledDT) println(metricsDT.accuracy)//0.6273062730627307 /*决策树的PR曲线和AOC曲线*/ val dtMetrics = Seq(model_DT).map{ model => val scoreAndLabels = data_test.map { point => val score = model.predict(point.features) (if (score > 0.5) 1.0 else 0.0, point.label) } val metrics = new BinaryClassificationMetrics(scoreAndLabels) (model.getClass.getSimpleName, metrics.areaUnderPR, metrics.areaUnderROC) } val allMetrics = dtMetrics allMetrics.foreach{ case (m, pr, roc) => println(f"$m, Area under PR: ${pr * 100.0}%2.4f%%, Area under ROC: ${roc * 100.0}%2.4f%%") } /* DecisionTreeModel, Area under PR: 74.2335%, Area under ROC: 62.4326% */
4,模型参数调优(可以调解长度和纯度两方面考虑)
4.1 构建调参函数
/*调参函数*/ def trainDTWithParams(input: RDD[LabeledPoint], maxDepth: Int, impurity: Impurity) = { DecisionTree.train(input, Algo.Classification, impurity, maxDepth) }
4.2 调解树的深度评估函数 (提高树的深度可以得到更精确的模型(这和预期一致,因为模型在更大的树深度下会变得更加复杂)。然而树的深度越大,模型对训练数据过拟合程度越严重)
/*改变深度*/ val dtResultsEntropy = Seq(1, 2, 3, 4, 5, 10, 20).map { param => val model = trainDTWithParams(data_train, param, Entropy) val scoreAndLabels = data_test.map { point => val score = model.predict(point.features) (if (score > 0.5) 1.0 else 0.0, point.label) } val metrics = new BinaryClassificationMetrics(scoreAndLabels) (s"$param tree depth", metrics.areaUnderROC) } dtResultsEntropy.foreach { case (param, auc) => println(f"$param, " + f"AUC = ${auc * 100}%2.2f%%") } /* 1 tree depth, AUC = 58.57% 2 tree depth, AUC = 60.69% 3 tree depth, AUC = 61.40% 4 tree depth, AUC = 61.30% 5 tree depth, AUC = 62.43% 10 tree depth, AUC = 62.26% 20 tree depth, AUC = 60.59% */
2,调解纯度参数 (差异不是很明显。。)
/*改变纯度*/ val dtResultsEntropy = Seq(Gini,Entropy).map { param => val model = trainDTWithParams(data_train, 5, param) val scoreAndLabels = data_test.map { point => val score = model.predict(point.features) (if (score > 0.5) 1.0 else 0.0, point.label) } val metrics = new BinaryClassificationMetrics(scoreAndLabels) (s"$param tree depth", metrics.areaUnderROC) } dtResultsEntropy.foreach { case (param, auc) => println(f"$param, " + f"AUC = ${auc * 100}%2.2f%%") } /* org.apache.spark.mllib.tree.impurity.Gini$@32d8e58d tree depth, AUC = 62.37% org.apache.spark.mllib.tree.impurity.Entropy$@1ddba7a0 tree depth, AUC = 62.43% */
(三) 交叉验证
1,数据集分类
创建三个数据集:训练集
评估集(类似上述测试集用于模型参数的调优,比如 lambda 和步长)
测试集(不用于模型的训练和参数调优,只用于估计模型在新数据中性能)
2,交叉验证的常用方法
一个流行的方法是 K- 折叠交叉验证,其中数据集被分成 K 个不重叠的部分。用数据中的 K-1 份训练模型,剩下一部分测试模型。而只分训练集和测试集可以看做是 2- 折叠交叉验证。
还有“留一交叉验证”和“随机采样”。更多资料详见 http://en.wikipedia.org/wiki/Cross-validation_(statistics) 。
标签:compute 分析 lan apach .class style vector stand nsf
原文地址:http://www.cnblogs.com/ksWorld/p/6884313.html