风控策略的自动化生成-利用决策树分分钟生成上千条策略
一、数据说明及读取
1、数据集信息
2、数据属性信息
3、读取数据
#数据读取
import pandas as pd
import numpy as np
pd.set_option('display.max_columns', None)#显示所有的列
path = '/Users/wuzhengxiang/Documents/DataSets/RizhaoGongJiJin/train.csv'
train = pd.read_csv(path).fillna(-1)
train.columns
Index(['id', 'XINGBIE', 'CSNY', 'HYZK', 'ZHIYE', 'ZHICHEN', 'ZHIWU', 'XUELI',
'DWJJLX', 'DWSSHY', 'GRJCJS', 'GRZHZT', 'GRZHYE', 'GRZHSNJZYE',
'GRZHDNGJYE', 'GRYJCE', 'DWYJCE', 'DKFFE', 'DKYE', 'DKLL', 'label'],
dtype='object')
train.head()#查看前面的数据
id XINGBIE CSNY HYZK ZHIYE ZHICHEN ZHIWU XUELI DWJJLX \
0 train_0 1 1038672000 90 90 999 0 99 150
1 train_1 2 504892800 90 90 999 0 99 110
2 train_2 1 736185600 90 90 999 0 99 150
3 train_3 1 428515200 90 90 999 0 99 150
4 train_4 2 544204800 90 90 999 0 99 900
DWSSHY GRJCJS GRZHZT GRZHYE GRZHSNJZYE GRZHDNGJYE GRYJCE \
0 12 1737.0 1 3223.515 801.310 837.000 312.00
1 0 4894.0 1 18055.195 53213.220 1065.200 795.84
2 9 10297.0 1 27426.600 13963.140 7230.020 1444.20
3 7 10071.5 1 111871.130 99701.265 2271.295 1417.14
4 14 2007.0 1 237.000 11028.875 35.780 325.50
DWYJCE DKFFE DKYE DKLL label
0 312.00 175237 154112.935 2.708 0
1 795.84 300237 298252.945 2.979 0
2 1444.20 150237 147339.130 2.708 0
3 1417.14 350237 300653.780 2.708 0
4 325.50 150237 145185.010 2.708 0
#构建训练集
X = train.loc[:,'XINGBIE':'DKLL']
train['label'] =
二、构建决策树
训练一个决策树,这里限制了最大深度和最小样本树 from sklearn import tree
tree.DecisionTreeClassifier( =
max_depth=3,
min_samples_leaf=50
)
clf.fit(X, Y) =
三、决策树的可视化
1、plot_tree(太丑,不推荐)
#包里自带的,有点丑
tree.plot_tree(clf)
plt.show()
2、graphviz对决策树进行可视化
import graphviz
tree.export_graphviz( =
out_file=None,
X.columns, =
class_names=['good','bad'],
filled=True, rounded=True,
special_characters=True)
graphviz.Source(dot_data) =
graph
3、dtreeviz对决策树进行可视化
from dtreeviz.trees import dtreeviz
X.iloc[77,:] =
dtreeviz(clf,X,Y, =
feature_names=np.array(X.columns),
class_names={0:'good',1:'bad'},
X = testX)
viz.view()
我们把树的深度再加深到5看看,树更复杂了 from sklearn import tree
clf = tree.DecisionTreeClassifier(
max_depth=5,
min_samples_leaf=50
)
clf.fit(X, Y) =
from dtreeviz.trees import dtreeviz
testX = X.iloc[77,:]
viz = dtreeviz(clf,X,Y,
feature_names=np.array(X.columns),
class_names={0:'good',1:'bad'},
X = testX)
viz.view()
viz = dtreeviz(clf,X,Y,
orientation ='LR', # left-right orientation
feature_names=np.array(X.columns),
class_names={0:'good',1:'bad'},
X = testX)
viz.view()
#如果只想可视化预测路径,则需要设置参数 show_just_path=True
viz = dtreeviz(clf,X,Y,
feature_names=np.array(X.columns),
class_names={0:'good',1:'bad'},
orientation ='LR', # left-right orientation
show_just_path=True,
X = testX)
viz.view()
四、决策规则提取
1、决策树的生成的结构探索
from sklearn.datasets import load_iris
from sklearn import tree
iris = load_iris()
clf = tree.DecisionTreeClassifier()
clf = clf.fit(iris.data, iris.target)
clf.classes_
[x for x in dir(clf) if not x.startswith('_')]
classes_:分类标签的取值,即y的唯一值集合 max_features_:最大特征数 n_classes_:类别数,如2分类或多分类等,即classes_属性中的长度 n_features_in_:输入特征数量,等价于老版sklearn中的n_features_,现已弃用,并推荐n_features_in_ n_outputs:多输出的个数,即决策树不仅可以用于实现单一的分类问题,还可同时实现多个分类问题,例如给定一组人物特征,用于同时判断其是男/女、胖/瘦和高矮,这是3个分类问题,即3输出(需要区别理解多分类和多输出任务) tree_:毫无疑问,这个tree_就是今天本文的重点,是在决策树训练之后新增的属性集,其中存储了决策树是如何存储的。
Array-based representation of a binary decision tree.
left_child:size类型(无符号整型),代表了当前节点的左子节点的索引 right_child:类似于left_child feature:size类型,代表了当前节点用于分裂的特征索引,即在训练集中用第几列特征进行分裂 threshold:double类型,代表了当前节点选用相应特征时的分裂阈值,一般是≤该阈值时进入左子节点,否则进入右子节点 n_node_samples:size类型,代表了训练时落入到该节点的样本总数。显然,父节点的n_node_samples将等于其左右子节点的n_node_samples之和。
# 决策树结构探索
dir(clf.tree_)
['apply','capacity', 'children_left','children_right',
'compute_feature_importances','compute_partial_dependence',
'decision_path','feature',
'impurity','max_depth',
'max_n_classes','n_classes',
'n_features','n_leaves',
'n_node_samples','n_outputs','node_count',
'predict','threshold',
'value', 'weighted_n_node_samples']
clf.tree_.children_left
clf.tree_.children_right
clf.tree_.feature
clf.tree_.capacity
clf.tree_.threshold
clf.tree_.value
clf.tree_.impurity
clf.tree_.decision_path
训练后的决策树共包含5个节点,其中3个叶子节点 通过children_left和children_right两个属性,可以知道第0个节点(也就是根节点)的左子节点索引为1,右子节点索引为2,;第1个节点的左右子节点均为-1,意味着该节点即为叶子节点;第2个节点的左右子节点分别为3和4,说明它是一个内部节点,并做了进一步分裂 通过feature和threshold两个属性,可以知道第0个节点(根节点)使用索引为3的特征(对应第4列特征)进行分裂,且其最优分割阈值为0.8;第1个节点因为是叶子节点,所以不再分裂,其对应feature和threshold字段均为-2 通过value属性,可以查看落入每个节点的各类样本数量,由于鸢尾花数据集是一个三分类问题,且该决策树共有5个节点,所以value的取值为一个5×3的二维数组,例如第一行代表落入根节点的样本计数为[50, 50, 50],第二行代表落入左子节点的样本计数为[50, 0, 0],由于已经是纯的了,所以不再继续分裂。 另外,tree中实际上并未直接标出各叶节点所对应的标签值,但完全可通过value属性来得到,即各叶子节点中落入样本最多的类别即为相应标签。甚至说,不仅可知道对应标签,还可通过计算数量之比得到相应的概率!
2、老方法提取决策树规则
数据读取 import pandas as pd
import numpy as np
pd.set_option('display.max_columns', None)#显示所有的列
path = '/Users/wuzhengxiang/Documents/DataSets/RizhaoGongJiJin/train.csv'
train = pd.read_csv(path).fillna(-1)
train.columns
构建训练集 X = train.loc[:,'XINGBIE':'DKLL']
Y = train['label']
训练一个决策树,这里限制了最大深度和最小样本树from sklearn import tree
clf = tree.DecisionTreeClassifier(
max_depth=3,
min_samples_leaf=50
)
clf = clf.fit(X, Y)
决策树规则提取-老方法
from sklearn.tree import _tree
def tree_to_code(tree, feature_names):
tree_ = tree.tree_
feature_name = [
feature_names[i] if i != _tree.TREE_UNDEFINED else "undefined!"
for i in tree_.feature
]
print ("def tree({}):".format(", ".join(feature_names)))
def recurse(node, depth):
indent = " " * depth
if tree_.feature[node] != _tree.TREE_UNDEFINED:
name = feature_name[node]
threshold = tree_.threshold[node]
print("{}if {} <= {}:".format(indent, name, threshold))
recurse(tree_.children_left[node], depth + 1)
print("{}else: # if {} > {}".format(indent, name, threshold))
recurse(tree_.children_right[node], depth + 1)
else:
print("{}return {}".format(indent, tree_.value[node]))
recurse(0, 1)
tree_to_code(clf,X.columns)
return [[161. 0.]]
3、新方法提取决策树规则
def XiaoWuGe_Get_Rules(clf,X):
n_nodes = clf.tree_.node_count
children_left = clf.tree_.children_left
children_right = clf.tree_.children_right
feature = clf.tree_.feature
threshold = clf.tree_.threshold
value = clf.tree_.value
node_depth = np.zeros(shape=n_nodes, dtype=np.int64)
is_leaves = np.zeros(shape=n_nodes, dtype=bool)
stack = [(0, 0)]
while len(stack) > 0:
node_id, depth = stack.pop()
node_depth[node_id] = depth
is_split_node = children_left[node_id] != children_right[node_id]
if is_split_node:
stack.append((children_left[node_id], depth+1))
stack.append((children_right[node_id], depth+1))
else:
is_leaves[node_id] = True
feature_name = [
X.columns[i] if i != _tree.TREE_UNDEFINED else "undefined!"
for i in clf.tree_.feature]
ways = []
depth = []
feat = []
nodes = []
rules = []
for i in range(n_nodes):
if is_leaves[i]:
while depth[-1] >= node_depth[i]:
depth.pop()
ways.pop()
feat.pop()
nodes.pop()
if children_left[i-1]==i:
a='{f}<={th}'.format(f=feat[-1],th=round(threshold[nodes[-1]],4))
ways[-1]=a
last =' & '.join(ways)+':'+str(value[i][0][0])+':'+str(value[i][0][1])
rules.append(last)
else:
a='{f}>{th}'.format(f=feat[-1],th=round(threshold[nodes[-1]],4))
ways[-1]=a
last = ' & '.join(ways)+':'+str(value[i][0][0])+':'+str(value[i][0][1])
rules.append(last)
else:
if i==0:
ways.append(round(threshold[i],4))
depth.append(node_depth[i])
feat.append(feature_name[i])
nodes.append(i)
else:
while depth[-1] >= node_depth[i]:
depth.pop()
ways.pop()
feat.pop()
nodes.pop()
if i==children_left[nodes[-1]]:
w='{f}<={th}'.format(f=feat[-1],th=round(threshold[nodes[-1]],4))
else:
w='{f}>{th}'.format(f=feat[-1],th=round(threshold[nodes[-1]],4))
ways[-1] = w
ways.append(round(threshold[i],4))
depth.append(node_depth[i])
feat.append(feature_name[i])
nodes.append(i)
return rules
4、利用函数对规则进行提取
#训练一个决策树,对规则进行提取
clf = tree.DecisionTreeClassifier(max_depth=10,
min_samples_leaf=50)
clf = clf.fit(X, Y)
Rules = XiaoWuGe_Get_Rules(clf,X)
Rules[0:5] # 查看前5条规则
['GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & DKYE<=67419.1094:45.0:8.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & DKYE >67419.1094:61.0:3.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE >663.54 & GRZHYE<=45622.4883 & DKYE<=1825.5625:63.0:2.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE >663.54 & GRZHYE<=45622.4883 & DKYE >1825.5625:188.0:0.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE >663.54 & GRZHYE >45622.4883:46.0:4.0']
len(Rules) # 查看规则总数
182
clf = tree.DecisionTreeClassifier(max_depth=15,min_samples_leaf=20)
clf = clf.fit(X, Y)
Rules = Get_Rules(clf,X)
Rules[0:5]
['GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & GRZHSNJZYE<=19428.9082 & DKFFE<=142737.0 & CSNY<=600926400.0:54.0:0.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & GRZHSNJZYE<=19428.9082 & DKFFE<=142737.0 & CSNY >600926400.0:18.0:2.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & GRZHSNJZYE<=19428.9082 & DKFFE >142737.0:19.0:4.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE<=663.54 & GRZHSNJZYE >19428.9082:15.0:5.0',
'GRZHZT<=1.5 & DWSSHY<=14.5 & DWJJLX<=177.0 & DWJJLX<=115.0 & DKYE<=111236.2852 & DWSSHY<=4.5 & DWYJCE >663.54 & GRZHYE<=73608.0156 & DKYE<=1825.5625 & GRZHSNJZYE<=9524.7949:21.0:2.0']
len(Rules)
521
#可以遍历所有的规则
for i in Rules:
print(i)
五、挖掘更多的风控策略
评论