如前所述,通过传进来的空节点(已经设置最大层数为3)构造初始左右节点
如果不是构造根节点,都要指定父节点papa(为神马没有mama )
父节点其实很关键,尤其在calc_output的时候,其实是先让祖先分类,分剩下的才轮到子孙继续划分
tree_node_left = tree_node;
tree_node_right = tree_node;
if(nargin > 4)
tree_node_left.parent = papa;
tree_node_right.parent = papa;
end
把传进来的变量改个名字
Distr = weights;
trainpat = dataset;
traintarg = labels;
tr_size = size(trainpat, 2);
T_MIN = zeros(3,size(trainpat,1));
d_min = 1;
d_max = size(trainpat,1);
其实CART最根本的分类思想说起来很简单,但实际操作起来还是很麻烦,主要有很多细节要考虑---这就是理解一个算法和真正具体实现的区别
这整个for循环就是遍历(或者说穷举!)每一维数据的所有可能取值
因为输入本身就是离散数据,所以直接用每个输入值当作可能的取值
值得一提的应该是里面的while loop,还考虑了若干个输入数据去相同取值的情况
根据每个数据已知i的分类结果(+1或-1),把相应的权重加起来
这里用的分类准则应该是“误分类不纯度”,不过不确定,实在懒得动脑筋去想想(离奇!)
按说Gini Index应该是最常用的
然后把如果按照这一维的数据的这个取值分类得到的错误值,序号和分类标签,一起保存到T_MIN中(不知多少人能看懂这句话。。。。。表达的很差)
for 循环结束
for d = d_min : d_max;
[DS, IX] = sort(trainpat(d,:));
TS = traintarg(IX);
DiS = Distr(IX);
lDS = length(DS);
vPos = 0 * TS;
vNeg = vPos;
i = 1;
j = 1;
while i <= lDS
k = 0;
while i + k <= lDS && DS(i) == DS(i+k)
if(TS(i+k) > 0)
vPos(j) = vPos(j) + DiS(i+k);
else
vNeg(j) = vNeg(j) + DiS(i+k);
end
k = k + 1;
end
i = i + k;
j = j + 1;
end
vNeg = vNeg(1:j-1);
vPos = vPos(1:j-1);
Error = zeros(1, j - 1);
InvError = Error;
IPos = vPos;
INeg = vNeg;
for i = 2 : length(IPos)
IPos(i) = IPos(i-1) + vPos(i);
INeg(i) = INeg(i-1) + vNeg(i);
end
Ntot = INeg(end);
Ptot = IPos(end);
for i = 1 : j - 1
Error(i) = IPos(i) + Ntot - INeg(i);
InvError(i) = INeg(i) + Ptot - IPos(i);
end
idx_of_err_min = find(Error == min(Error));
if(length(idx_of_err_min) < 1)
idx_of_err_min = 1;
end
if(length(idx_of_err_min) <1)
idx_of_err_min = idx_of_err_min;
end
idx_of_err_min = idx_of_err_min(1);
idx_of_inv_err_min = find(InvError == min(InvError));
if(length(idx_of_inv_err_min) < 1)
idx_of_inv_err_min = 1;
end
idx_of_inv_err_min = idx_of_inv_err_min(1);
if(Error(idx_of_err_min) < InvError(idx_of_inv_err_min))
T_MIN(1,d) = Error(idx_of_err_min);
T_MIN(2,d) = idx_of_err_min;
T_MIN(3,d) = -1;
else
T_MIN(1,d) = InvError(idx_of_inv_err_min);
T_MIN(2,d) = idx_of_inv_err_min;
T_MIN(3,d) = 1;
end
end
然后找出错误值最小的数据维度就是这个节点的划分维数,划分数值又做了点手脚:去该数据和后面数据的平均
这里才分完了第一层!!
前面说过,左节点右节点的划分值是一样的,也就是通过right_constraint和left_constraint哪个有值来判断该节点是左节点还是右节点
dim = [];
best_dim = find(T_MIN(1,:) == min(T_MIN(1,:)));
dim = best_dim(1);
tree_node_left.dim = dim;
tree_node_right.dim = dim;
TDS = sort(trainpat(dim,:));
lDS = length(TDS);
DS = TDS * 0;
i = 1;
j = 1;
while i <= lDS
k = 0;
while i + k <= lDS && TDS(i) == TDS(i+k)
DS(j) = TDS(i);
k = k + 1;
end
i = i + k;
j = j + 1;
end
DS = DS(1:j-1);
split = (DS(T_MIN(2,dim)) + DS(min(T_MIN(2,dim) + 1, length(DS)))) / 2;
split_error = T_MIN(1,dim);
tree_node_left.right_constrain = split;
tree_node_right.left_constrain = split;