public void Update(CARTNode node)
{
this.j = node.j;
this.splitVal = node.splitVal;
this.output = node.output;
this.region = node.region;
}
private static CARTTree GetSubTree(CARTTree tree)
{
var root = tree._root;
var stack = new Stack <CARTNode>();
AccessNonLeaf(root, stack);
double minAlpha = double.MaxValue; // 最小alpha
CARTNode minNode = null; // 最小alpha对应的内部节点
while (stack.Count > 0)
{
var t = stack.Pop(); // 某一内部节点
// 以 t 为根节点的子树的所有叶节点
var leaves_t = CARTUtil.GetLeaves(t);
// 以 t 为单节点，平方误差为
var var_t = CARTUtil.GetVar(t);
// 以 t 为根节点的子树，平方误差为
double var_subtree = 0;
for (int i = 0; i < leaves_t.Count; i++)
{
var leaf = leaves_t[i];
var_subtree += CARTUtil.GetVar(leaf);
}
var alpha_t = (var_t - var_subtree) / (leaves_t.Count - 1);
if (minAlpha > alpha_t)
{
minAlpha = alpha_t;
minNode = t;
}
}
// 获得最小的alpha，则对此节点剪枝，
// 需要注意的是，由于最终要生成一个子树序列，所以，我们不对原来的树剪枝，而是复制一个树并剪枝
return(PrunedClone(tree, minNode));
}
private static void AccessNonLeaf(CARTNode node, Stack <CARTNode> stack)
{
// node有子节点，且不是根结点，说明是内部叶节点
if (node.left != null && node.parent != null)
{
stack.Push(node);
AccessNonLeaf(node.left, stack);
AccessNonLeaf(node.right, stack);
}
}
/// <summary>
/// 获取方差，作为回归问题中的预测误差
/// </summary>
/// <param name="points"></param>
/// <returns></returns>
public static double GetVar(CARTNode node)
{
double ave = 0;
if (node.left != null)
{
ave = node.region.Sum(p => p.vals.LastOrDefault()) / node.region.Count;
}
else
{
ave = node.output;
}
return(node.region.Sum(p => Math.Pow(p.vals.LastOrDefault() - ave, 2)));
}
private static CARTTree PrunedClone(CARTTree tree, CARTNode node)
{
var queue = new Queue <CARTNode>(); // 原始树队列
queue.Enqueue(tree._root);
var root_1 = new CARTNode(); // 新树根节点
var queue_1 = new Queue <CARTNode>(); // 同步队列
queue_1.Enqueue(root_1);
while (queue.Count > 0)
{
var n = queue.Dequeue();
var n_1 = queue_1.Dequeue();
if (n == node) // 遇到需要被剪枝的内部节点，则需要剪枝为叶节点
{
// 设置叶节点的必要字段
n_1.output = n.region.Sum(p => p.vals.LastOrDefault()) / n.region.Count;
n_1.region = n.region;
}
else
{
n_1.Update(n); // 更新节点的固有属性（字段）
if (n.left != null)
{
// 非叶节点
n_1.left = new CARTNode()
{
parent = n_1
};
n_1.right = new CARTNode()
{
parent = n_1
};
queue.Enqueue(n.left);
queue.Enqueue(n.right);
queue_1.Enqueue(n_1.left);
queue_1.Enqueue(n_1.right);
}
// else，是叶节点，无需其他操作
}
}
return(new CARTTree()
{
_root = root_1
});
}
/// <summary>
/// 递归获取模型输出值
/// </summary>
/// <param name="point"></param>
/// <param name="node"></param>
/// <returns></returns>
private static double Judge(CARTPoint point, CARTNode node)
{
if (node.left != null)
{
return(node.output);
}
else
{
if (point.vals[node.j] < node.splitVal)
{
return(Judge(point, node.left));
}
else
{
return(Judge(point, node.right));
}
}
}
/// <summary>
/// 获取以指定节点为根结点的子树中的所有叶节点
/// </summary>
/// <param name="node"></param>
/// <returns></returns>
public static List <CARTNode> GetLeaves(CARTNode node)
{
var list = new List <CARTNode>();
var queue = new Queue <CARTNode>();
queue.Enqueue(node);
while (queue.Count > 0)
{
var n = queue.Dequeue();
if (n.left == null)
{
list.Add(n);
}
else
{
queue.Enqueue(n.left);
queue.Enqueue(n.right);
}
}
return(list);
}
/// <summary>
/// 生成决策树
/// </summary>
/// <param name="node"></param>
/// <param name="attrIdxs"></param>
/// <param name="points"></param>
private static void Create(CARTNode node, List <int> attrIdxs, List <CARTPoint> points)
{
node.region = points;
// 根据CART分裂策略，分裂后的区域内样本点数量至少为1，不可能为0
if (points.Count == 1)
{
// 如果为1，则不再分裂，直接设置为叶节点
node.output = points[0].vals.LastOrDefault();
}
else
{
var ave = points.Sum(p => p.vals.LastOrDefault()) / points.Count;
// 没有可用于分裂的属性，则设置叶节点
// 输出值的估计为区域中样本点输出值的均值
if (attrIdxs.Count == 0)
{
node.output = ave;
}
else
{
// 先计算整体的样本点的方差，如果小于阈值，则不分裂
double squareErr = 0;
foreach (var p in points)
{
squareErr += (p.vals.LastOrDefault() - ave) * (p.vals.LastOrDefault() - ave);
}
if (squareErr < ave / 1000)
{
// 如果方差小于一个阈值，则停止分裂，这里为了简单起见，阈值hardcode
node.output = ave;
}
else
{
TempResult minTemp = null; // 最小平方误差
int minJ = 0; // 对应的分裂属性索引
for (var i = 0; i < attrIdxs.Count; i++)
{
var j = attrIdxs[i]; // 输入属性的索引
var temp = CARTUtil.SquareError(j, points);
if (minTemp == null || temp.lossVal < minTemp.lossVal)
{
minTemp = temp;
minJ = j;
}
}
// 得到最小平方误差
node.j = minJ;
node.splitVal = minTemp.splitVal;
node.left = new CARTNode()
{
parent = node
};
node.right = new CARTNode()
{
parent = node
};
var leftAttrIdxs = attrIdxs.Where(idx => idx != minJ).Select(idx => idx).ToList();
var rightAttrIdxs = attrIdxs.Where(idx => idx != minJ).Select(idx => idx).ToList();
// 递归创建左右子节点
Create(node.left, leftAttrIdxs, minTemp.region1);
Create(node.right, rightAttrIdxs, minTemp.region2);
}
}
}
}
