机器学习---决策树之画图

sklearn机器学习—决策树之画图

--------仅用于个人学习知识整理和R语言/python代码整理
--------本文所使用数据全为脱敏模拟数据

---

1.前言

项目中会需要对决策树的树图进行修改，看起来更方便（相当于要对比每个节点1的概率，和基础的1的概率之间的lift提升），此处特别感谢某团初始代码思路提供，之后代码基于某团思路修改

2.二分类tree 增加lift及1 rate

0. 多分类的优雅？代码

感谢业务需求，遇到了一个7分类的树图，飞速改出灵活的代码

def dot_data_with_lift(model,dot_data):
        from functools  import reduce
        lift_list=[]
        model_value=list(model.tree_.value[0][0])
        base_ttl=reduce(lambda x,y:x+y,model_value)

        value_index=[]
        for i in range(0,len(model.tree_.value)):   
                ##要换掉整个这一层树的文本值
                model_value_i_raw=list(model.tree_.value[i][0])
                ##转int 才能和输出对上
                model_value_i=[int(c) for c in model_value_i_raw]
                if str(model_value_i) in value_index:
                    continue
                else:
                    ###一层的数据需要拼起来一起替换
                    lift_add=''
                    hit_add=''
                    for j in range(0,len(model_value)):
                        lift_base=model.tree_.value[0][0][j]/base_ttl
                        base_ttl_i=reduce(lambda x,y:x+y,model_value_i)
                        hit_rate=model.tree_.value[i][0][j]/base_ttl_i
                        lift=model.tree_.value[i][0][j]/base_ttl_i/lift_base
                        lift_add=lift_add+"<br/>lift "+str(j)+" = "+str(np.around(lift,decimals=4))+"   hitrate "+str(j)+" = "+str(np.around(hit_rate,decimals=4))
                    dot_data=dot_data.replace(str(model_value_i),str(model_value_i)+lift_add)
                    value_index.append(str(model_value_i))
        return dot_data

1. 画图code

核心思路是替换export_graphviz中，树的结构，加入新计算的lift和1的比例，以下改编自某团code

def dot_data_with_lift(model,dot_data):
    lift_list=[]
    lift_base=model.tree_.value[0][0][1]/(model.tree_.value[0][0][1]+model.tree_.value[0][0][0])
    value_index=[]
    for i in range(0,len(model.tree_.value)):   
        if str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1])]) in value_index:
            continue
        else:
### lift=当前节点1的概率 / 根节点1的概率
            lift=model.tree_.value[i][0][1]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0])/lift_base
### hit_rate=当前节点1的概率 
            hit_rate=model.tree_.value[i][0][1]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0])
### 把原来的dot_data中，替换加入lift 1_rate
            dot_data=dot_data.replace(str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1])]),\
                             str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1])])+\
                             "<br/>lift = "+str(lift)+\
                             "<br/>1_rate = "+str(hit_rate))
            value_index.append(str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1])]))
    return dot_data

### 调用
model=joblib.load(out_put)
dot_data =dot_data_with_lift(model,tree.export_graphviz(model, out_file=None, 
                         feature_names=variable_list,  
                         filled=True, rounded=True,  
                         special_characters=True) )
graph = pydotplus.graph_from_dot_data(dot_data)   
png_name='DecisionTree_r1.png'
graph.write_png(png_name)

2. 效果

本文所使用数据全为脱敏模拟数据
可以看到每个节点的lift和1的比例

2.多分类tree 增加lift及1 rate

def dot_data_with_lift(model,dot_data):
        lift_list=[]

        lift_base_1=model.tree_.value[0][0][1]/(model.tree_.value[0][0][1]+model.tree_.value[0][0][0]+model.tree_.value[0][0][2])
        lift_base_2=model.tree_.value[0][0][2]/(model.tree_.value[0][0][1]+model.tree_.value[0][0][0]+model.tree_.value[0][0][2])

        value_index=[]
        for i in range(0,len(model.tree_.value)):   
            if str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1]),int(model.tree_.value[i][0][2])]) in value_index:
                continue
            else:
                lift_1=model.tree_.value[i][0][1]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0]+model.tree_.value[i][0][2])/lift_base_1
                hit_rate_1=model.tree_.value[i][0][1]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0]+model.tree_.value[i][0][2])
                lift_2=model.tree_.value[i][0][2]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0]+model.tree_.value[i][0][2])/lift_base_2
                hit_rate_2=model.tree_.value[i][0][2]/(model.tree_.value[i][0][1]+model.tree_.value[i][0][0]+model.tree_.value[i][0][2])
                dot_data=dot_data.replace(str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1]),int(model.tree_.value[i][0][2])]),\
                                 str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1]),int(model.tree_.value[i][0][2])])+\
                                 "<br/>class1_lift = "+str(np.around(lift_1,decimals=4))+\
                                 "<br/>class1_1概率 = "+str(np.around(hit_rate_1,decimals=4))+\
                                 "<br/>class2_lift = "+str(np.around(lift_2,decimals=4))+\
                                 "<br/>class2_1概率 = "+str(np.around(hit_rate_2,decimals=4))
                                         )
                value_index.append(str([int(model.tree_.value[i][0][0]),int(model.tree_.value[i][0][1]),int(model.tree_.value[i][0][2])]))
        return dot_data