决策树系列(五)——CART
CART,又名分类回归树,是在ID3的基础上进行优化的决策树,学习CART记住以下几个关键点:
(1)CART既能是分类树,又能是分类树;
(2)当CART是分类树时,采用GINI值作为节点分裂的依据;当CART是回归树时,采用样本的最小方差作为节点分裂的依据;
(3)CART是一棵二叉树。
接下来将以一个实际的例子对CART进行介绍:
表1 原始数据表
看电视时间 |
婚姻情况 |
职业 |
年龄 |
3 |
未婚 |
学生 |
12 |
4 |
未婚 |
学生 |
18 |
2 |
已婚 |
老师 |
26 |
5 |
已婚 |
上班族 |
47 |
2.5 |
已婚 |
上班族 |
36 |
3.5 |
未婚 |
老师 |
29 |
4 |
已婚 |
学生 |
21 |
从以下的思路理解CART:
分类树?回归树?
分类树的作用是通过一个对象的特征来预测该对象所属的类别,而回归树的目的是根据一个对象的信息预测该对象的属性,并以数值表示。
CART既能是分类树,又能是决策树,如上表所示,如果我们想预测一个人是否已婚,那么构建的CART将是分类树;如果想预测一个人的年龄,那么构建的将是回归树。
分类树和回归树是怎么做决策的?假设我们构建了两棵决策树分别预测用户是否已婚和实际的年龄,如图1和图2所示:
图1 预测婚姻情况决策树 图2 预测年龄的决策树
图1表示一棵分类树,其叶子节点的输出结果为一个实际的类别,在这个例子里是婚姻的情况(已婚或者未婚),选择叶子节点中数量占比最大的类别作为输出的类别;
图2是一棵回归树,预测用户的实际年龄,是一个具体的输出值。怎样得到这个输出值?一般情况下选择使用中值、平均值或者众数进行表示,图2使用节点年龄数据的平均值作为输出值。
CART如何选择分裂的属性?
分裂的目的是为了能够让数据变纯,使决策树输出的结果更接近真实值。那么CART是如何评价节点的纯度呢?如果是分类树,CART采用GINI值衡量节点纯度;如果是回归树,采用样本方差衡量节点纯度。节点越不纯,节点分类或者预测的效果就越差。
GINI值的计算公式:
节点越不纯,GINI值越大。以二分类为例,如果节点的所有数据只有一个类别,则 ,如果两类数量相同,则
。
回归方差计算公式:
方差越大,表示该节点的数据越分散,预测的效果就越差。如果一个节点的所有数据都相同,那么方差就为0,此时可以很肯定得认为该节点的输出值;如果节点的数据相差很大,那么输出的值有很大的可能与实际值相差较大。
因此,无论是分类树还是回归树,CART都要选择使子节点的GINI值或者回归方差最小的属性作为分裂的方案。即最小化(分类树):
或者(回归树):
CART如何分裂成一棵二叉树?
节点的分裂分为两种情况,连续型的数据和离散型的数据。
CART对连续型属性的处理与C4.5差不多,通过最小化分裂后的GINI值或者样本方差寻找最优分割点,将节点一分为二,在这里不再叙述,详细请看C4.5。
对于离散型属性,理论上有多少个离散值就应该分裂成多少个节点。但CART是一棵二叉树,每一次分裂只会产生两个节点,怎么办呢?很简单,只要将其中一个离散值独立作为一个节点,其他的离散值生成另外一个节点即可。这种分裂方案有多少个离散值就有多少种划分的方法,举一个简单的例子:如果某离散属性一个有三个离散值X,Y,Z,则该属性的分裂方法有{X}、{Y,Z},{Y}、{X,Z},{Z}、{X,Y},分别计算每种划分方法的基尼值或者样本方差确定最优的方法。
以属性“职业”为例,一共有三个离散值,“学生”、“老师”、“上班族”。该属性有三种划分的方案,分别为{“学生”}、{“老师”、“上班族”},{“老师”}、{“学生”、“上班族”},{“上班族”}、{“学生”、“老师”},分别计算三种划分方案的子节点GINI值或者样本方差,选择最优的划分方法,如下图所示:
第一种划分方法:{“学生”}、{“老师”、“上班族”}
预测是否已婚(分类):
预测年龄(回归):
第二种划分方法:{“老师”}、{“学生”、“上班族”}
预测是否已婚(分类):
预测年龄(回归):
第三种划分方法:{“上班族”}、{“学生”、“老师”}
预测是否已婚(分类):
预测年龄(回归):
综上,如果想预测是否已婚,则选择{“上班族”}、{“学生”、“老师”}的划分方法,如果想预测年龄,则选择{“老师”}、{“学生”、“上班族”}的划分方法。
如何剪枝?
CART采用CCP(代价复杂度)剪枝方法。代价复杂度选择节点表面误差率增益值最小的非叶子节点,删除该非叶子节点的左右子节点,若有多个非叶子节点的表面误差率增益值相同小,则选择非叶子节点中子节点数最多的非叶子节点进行剪枝。
可描述如下:
令决策树的非叶子节点为。
a)计算所有非叶子节点的表面误差率增益值
b)选择表面误差率增益值最小的非叶子节点
(若多个非叶子节点具有相同小的表面误差率增益值,选择节点数最多的非叶子节点)。
c)对进行剪枝
表面误差率增益值的计算公式:
其中:
表示叶子节点的误差代价,
,
为节点的错误率,
为节点数据量的占比;
表示子树的误差代价,
,
为子节点i的错误率,
表示节点i的数据节点占比;
表示子树节点个数。
算例:
下图是其中一颗子树,设决策树的总数据量为40。
该子树的表面误差率增益值可以计算如下:
求出该子树的表面错误覆盖率为 ,只要求出其他子树的表面误差率增益值就可以对决策树进行剪枝。
程序实际以及源代码
流程图:
(1)数据处理
对原始的数据进行数字化处理,并以二维数据的形式存储,每一行表示一条记录,前n-1列表示属性,最后一列表示分类的标签。
如表1的数据可以转化为表2:
表2 初始化后的数据
看电视时间 |
婚姻情况 |
职业 |
年龄 |
3 |
未婚 |
学生 |
12 |
4 |
未婚 |
学生 |
18 |
2 |
已婚 |
老师 |
26 |
5 |
已婚 |
上班族 |
47 |
2.5 |
已婚 |
上班族 |
36 |
3.5 |
未婚 |
老师 |
29 |
4 |
已婚 |
学生 |
21 |
其中,对于“婚姻情况”属性,数字{1,2}分别表示{未婚,已婚 };对于“职业”属性{1,2,3, }分别表示{学生、老师、上班族};
代码如下所示:
static double[][] allData; //存储进行训练的数据
static List<String>[] featureValues; //离散属性对应的离散值
featureValues是链表数组,数组的长度为属性的个数,数组的每个元素为该属性的离散值链表。
(2)两个类:节点类和分裂信息
a)节点类Node
该类表示一个节点,属性包括节点选择的分裂属性、节点的输出类、孩子节点、深度等。注意,与ID3中相比,新增了两个属性:leafWrong和leafNode_Count分别表示叶子节点的总分类误差和叶子节点的个数,主要是为了方便剪枝。
树的节点
class Node
{
/// <summary>
/// 每一个节点的分裂值
/// </summary>
public List<String> features { get; set; }
/// <summary>
/// 分裂属性的类型{离散、连续}
/// </summary>
public String feature_Type { get; set; }
/// <summary>
/// 分裂属性的下标
/// </summary>
public String SplitFeature { get; set; }
//List<int> nums = new List<int>(); //行序号
/// <summary>
/// 每一个类对应的数目
/// </summary>
public double[] ClassCount { get; set; }
//int[] isUsed = new int[0]; //属性的使用情况 1:已用 2:未用
/// <summary>
/// 孩子节点
/// </summary>
public List<Node> childNodes { get; set; }
Node Parent = null;
/// <summary>
/// 该节点占比最大的类别
/// </summary>
public String finalResult { get; set; }
/// <summary>
/// 树的深度
/// </summary>
public int deep { get; set; }
/// <summary>
/// 最大的类下标
/// </summary>
public int result { get; set; }
/// <summary>
/// 子节点误差
/// </summary>
public int leafWrong { get; set; }
/// <summary>
/// 子节点数目
/// </summary>
public int leafNode_Count { get; set; }
/// <summary>
/// 数据量
/// </summary>
public int rowCount { get; set; }
public void setClassCount(double[] count)
{
this.ClassCount = count;
double max = ClassCount[0];
int result = 0;
for (int i = 1; i < ClassCount.Length; i++)
{
if (max < ClassCount[i])
{
max = ClassCount[i];
result = i;
}
}
this.result = result;
}
public double getErrorCount()
{
return rowCount - ClassCount[result];
}
}
树的节点
b)分裂信息类,该类存储节点进行分裂的信息,包括各个子节点的行坐标、子节点各个类的数目、该节点分裂的属性、属性的类型等。
分裂信息
class SplitInfo
{
/// <summary>
/// 分裂的属性下标
/// </summary>
public int splitIndex { get; set; }
/// <summary>
/// 数据类型
/// </summary>
public int type { get; set; }
/// <summary>
/// 分裂属性的取值
/// </summary>
public List<String> features { get; set; }
/// <summary>
/// 各个节点的行坐标链表
/// </summary>
public List<int>[] temp { get; set; }
/// <summary>
/// 每个节点各类的数目
/// </summary>
public double[][] class_Count { get; set; }
}
分裂信息
主方法findBestSplit(Node node,List<int> nums,int[] isUsed),该方法对节点进行分裂
其中:
node表示即将进行分裂的节点;
nums表示节点数据对一个的行坐标列表;
isUsed表示到该节点位置所有属性的使用情况;
findBestSplit的这个方法主要有以下几个组成部分:
1)节点分裂停止的判定
节点分裂条件如上文所述,源代码如下:
停止分裂的条件
public static bool ifEnd(Node node, double shang,int[] isUsed)
{
try
{
double[] count = node.ClassCount;
int rowCount = node.rowCount;
int maxResult = 0;
double maxRate = 0;
#region 数达到某一深度
int deep = node.deep;
if (deep >= 10)
{
maxResult = node.result + 1;
node.feature_Type="result";
node.features=new List<String>() { maxResult + ""
};
node.leafWrong=rowCount - Convert.ToInt32(count[maxResult-1]);
node.leafNode_Count=1;
return true;
}
#endregion
#region 纯度(其实跟后面的有点重了,记得要修改)
//maxResult = 1;
//for (int i = 1; i < count.Length; i++)
//{
// if (count[i] / rowCount >= 0.95)
// {
// node.feature_Type="result";
// node.features=new List<String> { "" + (i +
1) };
// node.leafNode_Count=1;
// node.leafWrong=rowCount - Convert.ToInt32
(count[i]);
// return true;
// }
//}
#endregion
#region 熵为0
if (shang == 0)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type="result";
node.features=new List<String> { maxResult + ""
};
node.leafWrong=rowCount - Convert.ToInt32(count
[maxResult - 1]);
node.leafNode_Count=1;
return true;
}
#endregion
#region 属性已经分完
//int[] isUsed = node.getUsed();
bool flag = true;
for (int i = 0; i < isUsed.Length - 1; i++)
{
if (isUsed[i] == 0)
{
flag = false;
break;
}
}
if (flag)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type=("result");
node.features=(new List<String> { "" +
(maxResult) });
node.leafWrong=(rowCount - Convert.ToInt32(count
[maxResult - 1]));
node.leafNode_Count=(1);
return true;
}
#endregion
#region 几点数少于100
if (rowCount < Limit_Node)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type="result";
node.features=new List<String> { "" + (maxResult)
};
node.leafWrong=rowCount - Convert.ToInt32(count
[maxResult - 1]);
node.leafNode_Count=1;
return true;
}
#endregion
return false;
}
catch (Exception e)
{
return false;
}
}
停止分裂的条件
2)寻找最优的分裂属性
寻找最优的分裂属性需要计算每一个分裂属性分裂后的GINI值或者样本方差,计算公式上文已给出,其中GINI值的计算代码如下:
GINI值计算
public static double getGini(double[] counts, int countAll)
{
double Gini = 1;
for (int i = 0; i < counts.Length; i++)
{
Gini = Gini - Math.Pow(counts[i] / countAll, 2);
}
return Gini;
}
GINI值计算
3)进行分裂,同时对子节点进行迭代处理
其实就是递归的过程,对每一个子节点执行findBestSplit方法进行分裂。
findBestSplit源代码:
节点选择属性和分裂
public static Node findBestSplit(Node node,List<int> nums,int[] isUsed)
{
try
{
//判断是否继续分裂
double totalShang = getGini(node.ClassCount, node.rowCount);
if (ifEnd(node, totalShang, isUsed))
{
return node;
}
#region 变量声明
SplitInfo info = new SplitInfo();
info.initial();
int RowCount = nums.Count; //样本总数
double jubuMax = 1; //局部最大熵
int splitPoint = 0; //分裂的点
double splitValue = 0; //分裂的值
#endregion
for (int i = 0; i < isUsed.Length - 1; i++)
{
if (isUsed[i] == 1)
{
continue;
}
#region 离散变量
if (type[i] == 0)
{
double[][] allCount = new double[allNum[i]][];
for (int j = 0; j < allCount.Length; j++)
{
allCount[j] = new double[classCount];
}
int[] countAllFeature = new int[allNum[i]];
List<int>[] temp = new List<int>[allNum[i]];
double[] allClassCount = node.ClassCount; //所有类别的数量
for (int j = 0; j < temp.Length; j++)
{
temp[j] = new List<int>();
}
for (int j = 0; j < nums.Count; j++)
{
int index = Convert.ToInt32(allData[nums[j]][i]);
temp[index - 1].Add(nums[j]);
countAllFeature[index - 1]++;
allCount[index - 1][Convert.ToInt32(allData[nums[j]][lieshu - 1]) - 1]++;
}
double allShang = 1;
int choose = 0;
double[][] jubuCount = new double[2][];
for (int k = 0; k < allCount.Length; k++)
{
if (temp[k].Count == 0)
continue;
double JubuShang = 0;
double[][] tempCount = new double[2][];
tempCount[0] = allCount[k];
tempCount[1] = new double[allCount[0].Length];
for (int j = 0; j < tempCount[1].Length; j++)
{
tempCount[1][j] = allClassCount[j] - allCount[k][j];
}
JubuShang = JubuShang + getGini(tempCount[0], countAllFeature[k]) * countAllFeature[k] / RowCount;
int nodecount = RowCount - countAllFeature[k];
JubuShang = JubuShang + getGini(tempCount[1], nodecount) * nodecount / RowCount;
if (JubuShang < allShang)
{
allShang = JubuShang;
jubuCount = tempCount;
choose = k;
}
}
if (allShang < jubuMax)
{
info.type = 0;
jubuMax = allShang;
info.class_Count = jubuCount;
info.temp[0] = temp[choose];
info.temp[1] = new List<int>();
info.features = new List<string>();
info.features.Add((choose + 1) + "");
info.features.Add("");
for (int j = 0; j < temp.Length; j++)
{
if (j == choose)
continue;
for (int k = 0; k < temp[j].Count; k++)
{
info.temp[1].Add(temp[j][k]);
}
if (temp[j].Count != 0)
{
info.features[1] = info.features[1] + (j + 1) + ",";
}
}
info.splitIndex = i;
}
}
#endregion
#region 连续变量
else
{
double[] leftCunt = new double[classCount];
//做节点各个类别的数量
double[] rightCount = new double[classCount];
//右节点各个类别的数量
double[] count1 = new double[classCount];
//子集1的统计量
double[] count2 = new double
[node.ClassCount.Length]; //子集2的统计量
for (int j = 0; j < node.ClassCount.Length;
j++)
{
count2[j] = node.ClassCount[j];
}
int all1 = 0;
//子集1的样本量
int all2 = nums.Count;
//子集2的样本量
double lastValue = 0;
//上一个记录的类别
double currentValue = 0;
//当前类别
double lastPoint = 0;
//上一个点的值
double currentPoint = 0;
//当前点的值
double[] values = new double[nums.Count];
for (int j = 0; j < values.Length; j++)
{
values[j] = allData[nums[j]][i];
}
QSort(values, nums, 0, nums.Count - 1);
double lianxuMax = 1;
//连续型属性的最大熵
#region 寻找最佳的分割点
for (int j = 0; j < nums.Count - 1; j++)
{
currentValue = allData[nums[j]][lieshu -
1];
currentPoint = (allData[nums[j]][i]);
if (j == 0)
{
lastValue = currentValue;
lastPoint = currentPoint;
}
if (currentValue != lastValue &&
currentPoint != lastPoint)
{
double shang1 = getGini(count1,
all1);
double shang2 = getGini(count2,
all2);
double allShang = shang1 * all1 /
(all1 + all2) + shang2 * all2 / (all1 + all2);
//allShang = (totalShang - allShang);
if (lianxuMax > allShang)
{
lianxuMax = allShang;
for (int k = 0; k <
count1.Length; k++)
{
leftCunt[k] = count1[k];
rightCount[k] = count2[k];
}
splitPoint = j;
splitValue = (currentPoint +
lastPoint) / 2;
}
}
all1++;
count1[Convert.ToInt32(currentValue) -
1]++;
count2[Convert.ToInt32(currentValue) -
1]--;
all2--;
lastValue = currentValue;
lastPoint = currentPoint;
}
#endregion
#region 如果超过了局部值,重设
if (lianxuMax < jubuMax)
{
info.type = 1;
info.splitIndex = i;
info.features=new List<string>()
{splitValue+""};
//finalPoint = splitPoint;
jubuMax = lianxuMax;
info.temp[0] = new List<int>();
info.temp[1] = new List<int>();
for (int k = 0; k < splitPoint; k++)
{
info.temp[0].Add(nums[k]);
}
for (int k = splitPoint; k < nums.Count;
k++)
{
info.temp[1].Add(nums[k]);
}
info.class_Count[0] = new double
[leftCunt.Length];
info.class_Count[1] = new double
[leftCunt.Length];
for (int k = 0; k < leftCunt.Length; k++)
{
info.class_Count[0][k] = leftCunt[k];
info.class_Count[1][k] = rightCount
[k];
}
}
#endregion
}
#endregion
}
#region 没有寻找到最佳的分裂点,则设置为叶节点
if (info.splitIndex == -1)
{
double[] finalCount = node.ClassCount;
double max = finalCount[0];
int result = 1;
for (int i = 1; i < finalCount.Length; i++)
{
if (finalCount[i] > max)
{
max = finalCount[i];
result = (i + 1);
}
}
node.feature_Type="result";
node.features=new List<String> { "" + result };
return node;
}
#endregion
#region 分裂
int deep = node.deep;
node.SplitFeature = ("" + info.splitIndex);
List<Node> childNode = new List<Node>();
int[][] used = new int[2][];
used[0] = new int[isUsed.Length];
used[1] = new int[isUsed.Length];
for (int i = 0; i < isUsed.Length; i++)
{
used[0][i] = isUsed[i];
used[1][i] = isUsed[i];
}
if (info.type == 0)
{
used[0][info.splitIndex] = 1;
node.feature_Type = ("离散");
}
else
{
//used[info.splitIndex] = 0;
node.feature_Type = ("连续");
}
List<int>[] rowIndex = info.temp;
List<String> features = info.features;
Node node1 = new Node();
Node node2 = new Node();
node1.setClassCount(info.class_Count[0]);
node2.setClassCount(info.class_Count[1]);
node1.rowCount = info.temp[0].Count;
node2.rowCount = info.temp[1].Count;
node1.deep = deep + 1;
node2.deep = deep + 1;
node1 = findBestSplit(node1, info.temp[0],used[0]);
node2 = findBestSplit(node2, info.temp[1], used[1]);
node.leafNode_Count = (node1.leafNode_Count
+node2.leafNode_Count);
node.leafWrong = (node1.leafWrong+node2.leafWrong);
node.features = (features);
childNode.Add(node1);
childNode.Add(node2);
node.childNodes = childNode;
#endregion
return node;
}
catch (Exception e)
{
Console.WriteLine(e.StackTrace);
return node;
}
}
节点选择属性和分裂
(4)剪枝
代价复杂度剪枝方法(CCP):
CCP代价复杂度剪枝
public static void getSeries(Node node)
{
Stack<Node> nodeStack = new Stack<Node>();
if (node != null)
{
nodeStack.Push(node);
}
if (node.feature_Type == "result")
return;
List<Node> childs = node.childNodes;
for (int i = 0; i < childs.Count; i++)
{
getSeries(node);
}
}
CCP代价复杂度剪枝
CART全部核心代码:
CART核心代码
/// <summary>
/// 判断是否还需要分裂
/// </summary>
/// <param name="node"></param>
/// <returns></returns>
public static bool ifEnd(Node node, double shang,int[] isUsed)
{
try
{
double[] count = node.ClassCount;
int rowCount = node.rowCount;
int maxResult = 0;
double maxRate = 0;
#region 数达到某一深度
int deep = node.deep;
if (deep >= 10)
{
maxResult = node.result + 1;
node.feature_Type="result";
node.features=new List<String>() { maxResult + ""
};
node.leafWrong=rowCount - Convert.ToInt32(count[maxResult-1]);
node.leafNode_Count=1;
return true;
}
#endregion
#region 纯度(其实跟后面的有点重了,记得要修改)
//maxResult = 1;
//for (int i = 1; i < count.Length; i++)
//{
// if (count[i] / rowCount >= 0.95)
// {
// node.feature_Type="result";
// node.features=new List<String> { "" + (i +
1) };
// node.leafNode_Count=1;
// node.leafWrong=rowCount - Convert.ToInt32
(count[i]);
// return true;
// }
//}
#endregion
#region 熵为0
if (shang == 0)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type="result";
node.features=new List<String> { maxResult + ""
};
node.leafWrong=rowCount - Convert.ToInt32(count
[maxResult - 1]);
node.leafNode_Count=1;
return true;
}
#endregion
#region 属性已经分完
//int[] isUsed = node.getUsed();
bool flag = true;
for (int i = 0; i < isUsed.Length - 1; i++)
{
if (isUsed[i] == 0)
{
flag = false;
break;
}
}
if (flag)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type=("result");
node.features=(new List<String> { "" +
(maxResult) });
node.leafWrong=(rowCount - Convert.ToInt32(count
[maxResult - 1]));
node.leafNode_Count=(1);
return true;
}
#endregion
#region 几点数少于100
if (rowCount < Limit_Node)
{
maxRate = count[0] / rowCount;
maxResult = 1;
for (int i = 1; i < count.Length; i++)
{
if (count[i] / rowCount >= maxRate)
{
maxRate = count[i] / rowCount;
maxResult = i + 1;
}
}
node.feature_Type="result";
node.features=new List<String> { "" + (maxResult)
};
node.leafWrong=rowCount - Convert.ToInt32(count
[maxResult - 1]);
node.leafNode_Count=1;
return true;
}
#endregion
return false;
}
catch (Exception e)
{
return false;
}
}
#region 排序算法
public static void InsertSort(double[] values, List<int> arr,
int StartIndex, int endIndex)
{
for (int i = StartIndex + 1; i <= endIndex; i++)
{
int key = arr[i];
double init = values[i];
int j = i - 1;
while (j >= StartIndex && values[j] > init)
{
arr[j + 1] = arr[j];
values[j + 1] = values[j];
j--;
}
arr[j + 1] = key;
values[j + 1] = init;
}
}
static int SelectPivotMedianOfThree(double[] values, List<int> arr, int low, int high)
{
int mid = low + ((high - low) >> 1);//计算数组中间的元素的下标
//使用三数取中法选择枢轴
if (values[mid] > values[high])//目标: arr[mid] <= arr[high]
{
swap(values, arr, mid, high);
}
if (values[low] > values[high])//目标: arr[low] <= arr[high]
{
swap(values, arr, low, high);
}
if (values[mid] > values[low]) //目标: arr[low] >= arr[mid]
{
swap(values, arr, mid, low);
}
//此时,arr[mid] <= arr[low] <= arr[high]
return low;
//low的位置上保存这三个位置中间的值
//分割时可以直接使用low位置的元素作为枢轴,而不用改变分割函数了
}
static void swap(double[] values, List<int> arr, int t1, int t2)
{
double temp = values[t1];
values[t1] = values[t2];
values[t2] = temp;
int key = arr[t1];
arr[t1] = arr[t2];
arr[t2] = key;
}
static void QSort(double[] values, List<int> arr, int low, int high)
{
int first = low;
int last = high;
int left = low;
int right = high;
int leftLen = 0;
int rightLen = 0;
if (high - low + 1 < 10)
{
InsertSort(values, arr, low, high);
return;
}
//一次分割
int key = SelectPivotMedianOfThree(values, arr, low,
high);//使用三数取中法选择枢轴
double inti = values[key];
int currentKey = arr[key];
while (low < high)
{
while (high > low && values[high] >= inti)
{
if (values[high] == inti)//处理相等元素
{
swap(values, arr, right, high);
right--;
rightLen++;
}
high--;
}
arr[low] = arr[high];
values[low] = values[high];
while (high > low && values[low] <= inti)
{
if (values[low] == inti)
{
swap(values, arr, left, low);
left++;
leftLen++;
}
low++;
}
arr[high] = arr[low];
values[high] = values[low];
}
arr[low] = currentKey;
values[low] = values[key];
//一次快排结束
//把与枢轴key相同的元素移到枢轴最终位置周围
int i = low - 1;
int j = first;
while (j < left && values[i] != inti)
{
swap(values, arr, i, j);
i--;
j++;
}
i = low + 1;
j = last;
while (j > right && values[i] != inti)
{
swap(values, arr, i, j);
i++;
j--;
}
QSort(values, arr, first, low - 1 - leftLen);
QSort(values, arr, low + 1 + rightLen, last);
}
#endregion
/// <summary>
/// 寻找最佳的分裂点
/// </summary>
/// <param name="num"></param>
/// <param name="node"></param>
public static Node findBestSplit(Node node,List<int> nums,int[] isUsed)
{
try
{
//判断是否继续分裂
double totalShang = getGini(node.ClassCount, node.rowCount);
if (ifEnd(node, totalShang, isUsed))
{
return node;
}
#region 变量声明
SplitInfo info = new SplitInfo();
info.initial();
int RowCount = nums.Count; //样本总数
double jubuMax = 1; //局部最大熵
int splitPoint = 0; //分裂的点
double splitValue = 0; //分裂的值
#endregion
for (int i = 0; i < isUsed.Length - 1; i++)
{
if (isUsed[i] == 1)
{
continue;
}
#region 离散变量
if (type[i] == 0)
{
double[][] allCount = new double[allNum[i]][];
for (int j = 0; j < allCount.Length; j++)
{
allCount[j] = new double[classCount];
}
int[] countAllFeature = new int[allNum[i]];
List<int>[] temp = new List<int>[allNum[i]];
double[] allClassCount = node.ClassCount; //所有类别的数量
for (int j = 0; j < temp.Length; j++)
{
temp[j] = new List<int>();
}
for (int j = 0; j < nums.Count; j++)
{
int index = Convert.ToInt32(allData[nums[j]][i]);
temp[index - 1].Add(nums[j]);
countAllFeature[index - 1]++;
allCount[index - 1][Convert.ToInt32(allData[nums[j]][lieshu - 1]) - 1]++;
}
double allShang = 1;
int choose = 0;
double[][] jubuCount = new double[2][];
for (int k = 0; k < allCount.Length; k++)
{
if (temp[k].Count == 0)
continue;
double JubuShang = 0;
double[][] tempCount = new double[2][];
tempCount[0] = allCount[k];
tempCount[1] = new double[allCount[0].Length];
for (int j = 0; j < tempCount[1].Length; j++)
{
tempCount[1][j] = allClassCount[j] - allCount[k][j];
}
JubuShang = JubuShang + getGini(tempCount[0], countAllFeature[k]) * countAllFeature[k] / RowCount;
int nodecount = RowCount - countAllFeature[k];
JubuShang = JubuShang + getGini(tempCount[1], nodecount) * nodecount / RowCount;
if (JubuShang < allShang)
{
allShang = JubuShang;
jubuCount = tempCount;
choose = k;
}
}
if (allShang < jubuMax)
{
info.type = 0;
jubuMax = allShang;
info.class_Count = jubuCount;
info.temp[0] = temp[choose];
info.temp[1] = new List<int>();
info.features = new List<string>();
info.features.Add((choose + 1) + "");
info.features.Add("");
for (int j = 0; j < temp.Length; j++)
{
if (j == choose)
continue;
for (int k = 0; k < temp[j].Count; k++)
{
info.temp[1].Add(temp[j][k]);
}
if (temp[j].Count != 0)
{
info.features[1] = info.features[1] + (j + 1) + ",";
}
}
info.splitIndex = i;
}
}
#endregion
#region 连续变量
else
{
double[] leftCunt = new double[classCount];
//做节点各个类别的数量
double[] rightCount = new double[classCount];
//右节点各个类别的数量
double[] count1 = new double[classCount];
//子集1的统计量
double[] count2 = new double
[node.ClassCount.Length]; //子集2的统计量
for (int j = 0; j < node.ClassCount.Length;
j++)
{
count2[j] = node.ClassCount[j];
}
int all1 = 0;
//子集1的样本量
int all2 = nums.Count;
//子集2的样本量
double lastValue = 0;
//上一个记录的类别
double currentValue = 0;
//当前类别
double lastPoint = 0;
//上一个点的值
double currentPoint = 0;
//当前点的值
double[] values = new double[nums.Count];
for (int j = 0; j < values.Length; j++)
{
values[j] = allData[nums[j]][i];
}
QSort(values, nums, 0, nums.Count - 1);
double lianxuMax = 1;
//连续型属性的最大熵
#region 寻找最佳的分割点
for (int j = 0; j < nums.Count - 1; j++)
{
currentValue = allData[nums[j]][lieshu -
1];
currentPoint = (allData[nums[j]][i]);
if (j == 0)
{
lastValue = currentValue;
lastPoint = currentPoint;
}
if (currentValue != lastValue &&
currentPoint != lastPoint)
{
double shang1 = getGini(count1,
all1);
double shang2 = getGini(count2,
all2);
double allShang = shang1 * all1 /
(all1 + all2) + shang2 * all2 / (all1 + all2);
//allShang = (totalShang - allShang);
if (lianxuMax > allShang)
{
lianxuMax = allShang;
for (int k = 0; k <
count1.Length; k++)
{
leftCunt[k] = count1[k];
rightCount[k] = count2[k];
}
splitPoint = j;
splitValue = (currentPoint +
lastPoint) / 2;
}
}
all1++;
count1[Convert.ToInt32(currentValue) -
1]++;
count2[Convert.ToInt32(currentValue) -
1]--;
all2--;
lastValue = currentValue;
lastPoint = currentPoint;
}
#endregion
#region 如果超过了局部值,重设
if (lianxuMax < jubuMax)
{
info.type = 1;
info.splitIndex = i;
info.features=new List<string>()
{splitValue+""};
//finalPoint = splitPoint;
jubuMax = lianxuMax;
info.temp[0] = new List<int>();
info.temp[1] = new List<int>();
for (int k = 0; k < splitPoint; k++)
{
info.temp[0].Add(nums[k]);
}
for (int k = splitPoint; k < nums.Count;
k++)
{
info.temp[1].Add(nums[k]);
}
info.class_Count[0] = new double
[leftCunt.Length];
info.class_Count[1] = new double
[leftCunt.Length];
for (int k = 0; k < leftCunt.Length; k++)
{
info.class_Count[0][k] = leftCunt[k];
info.class_Count[1][k] = rightCount
[k];
}
}
#endregion
}
#endregion
}
#region 没有寻找到最佳的分裂点,则设置为叶节点
if (info.splitIndex == -1)
{
double[] finalCount = node.ClassCount;
double max = finalCount[0];
int result = 1;
for (int i = 1; i < finalCount.Length; i++)
{
if (finalCount[i] > max)
{
max = finalCount[i];
result = (i + 1);
}
}
node.feature_Type="result";
node.features=new List<String> { "" + result };
return node;
}
#endregion
#region 分裂
int deep = node.deep;
node.SplitFeature = ("" + info.splitIndex);
List<Node> childNode = new List<Node>();
int[][] used = new int[2][];
used[0] = new int[isUsed.Length];
used[1] = new int[isUsed.Length];
for (int i = 0; i < isUsed.Length; i++)
{
used[0][i] = isUsed[i];
used[1][i] = isUsed[i];
}
if (info.type == 0)
{
used[0][info.splitIndex] = 1;
node.feature_Type = ("离散");
}
else
{
//used[info.splitIndex] = 0;
node.feature_Type = ("连续");
}
List<int>[] rowIndex = info.temp;
List<String> features = info.features;
Node node1 = new Node();
Node node2 = new Node();
node1.setClassCount(info.class_Count[0]);
node2.setClassCount(info.class_Count[1]);
node1.rowCount = info.temp[0].Count;
node2.rowCount = info.temp[1].Count;
node1.deep = deep + 1;
node2.deep = deep + 1;
node1 = findBestSplit(node1, info.temp[0],used[0]);
node2 = findBestSplit(node2, info.temp[1], used[1]);
node.leafNode_Count = (node1.leafNode_Count
+node2.leafNode_Count);
node.leafWrong = (node1.leafWrong+node2.leafWrong);
node.features = (features);
childNode.Add(node1);
childNode.Add(node2);
node.childNodes = childNode;
#endregion
return node;
}
catch (Exception e)
{
Console.WriteLine(e.StackTrace);
return node;
}
}
/// <summary>
/// GINI值
/// </summary>
/// <param name="counts"></param>
/// <param name="countAll"></param>
/// <returns></returns>
public static double getGini(double[] counts, int countAll)
{
double Gini = 1;
for (int i = 0; i < counts.Length; i++)
{
Gini = Gini - Math.Pow(counts[i] / countAll, 2);
}
return Gini;
}
#region CCP剪枝
public static void getSeries(Node node)
{
Stack<Node> nodeStack = new Stack<Node>();
if (node != null)
{
nodeStack.Push(node);
}
if (node.feature_Type == "result")
return;
List<Node> childs = node.childNodes;
for (int i = 0; i < childs.Count; i++)
{
getSeries(node);
}
}
/// <summary>
/// 遍历剪枝
/// </summary>
/// <param name="node"></param>
public static Node getNode1(Node node, Node nodeCut)
{
//List<Node> childNodes = node.getChild();
//double min = 100000;
////Node nodeCut = new Node();
//double temp = 0;
//for (int i = 0; i < childNodes.Count; i++)
//{
// if (childNodes[i].getType() != "result")
// {
// //if (!cutTree(childNodes[i]))
// temp = min;
// min = cutTree(childNodes[i], min);
// if (min < temp)
// nodeCut = childNodes[i];
// getNode1(childNodes[i], nodeCut);
// }
//}
//node.setChildNode(childNodes);
return null;
}
/// <summary>
/// 对每一个节点剪枝
/// </summary>
public static double cutTree(Node node, double minA)
{
int rowCount = node.rowCount;
double leaf = node.getErrorCount();
double[] values = getError1(node, 0, 0);
double treeWrong = values[0];
double son = values[1];
double rate = (leaf - treeWrong) / (son - 1);
if (minA > rate)
minA = rate;
//double var = Math.Sqrt(treeWrong * (1 - treeWrong /
rowCount));
//double panbie = treeWrong + var - leaf;
//if (panbie > 0)
//{
// node.setFeatureType("result");
// node.setChildNode(null);
// int result = (node.getResult() + 1);
// node.setFeatures(new List<String>() { "" + result
});
// //return true;
//}
return minA;
}
/// <summary>
/// 获得子树的错误个数
/// </summary>
/// <param name="node"></param>
/// <returns></returns>
public static double[] getError1(Node node, double treeError,
double son)
{
if (node.feature_Type == "result")
{
double error = node.getErrorCount();
son++;
return new double[] { treeError + error, son };
}
List<Node> childNode = node.childNodes;
for (int i = 0; i < childNode.Count; i++)
{
double[] values = getError1(childNode[i], treeError,
son);
treeError = values[0];
son = values[1];
}
return new double[] { treeError, son };
}
#endregion
CART核心代码
总结:
(1)CART是一棵二叉树,每一次分裂会产生两个子节点,对于连续性的数据,直接采用与C4.5相似的处理方法,对于离散型数据,选择最优的两种离散值组合方法。
(2)CART既能是分类数,又能是二叉树。如果是分类树,将选择能够最小化分裂后节点GINI值的分裂属性;如果是回归树,选择能够最小化两个节点样本方差的分裂属性。
(3)CART跟C4.5一样,需要进行剪枝,采用CCP(代价复杂度的剪枝方法)。
分类: 机器学习算法—从原理到实现