使用者流失預測

資料詳細情況請檢視：使用者流失分析

資料預處理

轉換資料類型和處理缺失值

#轉換資料類型
#TotalCharges總費用 和 MonthlyCharges每月費用 ，都應該是數值型資料，轉換TotalCharges為數值型資料
data['TotalCharges'] = pd.to_numeric(data['TotalCharges'],errors='coerce')
           

#處理缺失值
#用MonthlyCharges 的值 填充 TotalCharges的缺失值
data['TotalCharges'] = data['TotalCharges'].fillna(data['MonthlyCharges'])
           

特征工程

資料歸一化并删除不需要标簽

#删除使用者ID标簽
data =  data.iloc[:,1:]
           

#檢視非數值資料包含的特征資訊
def uni(label):
    print(label,'-----',data[label].unique())
    
data_object = data.select_dtypes(['object'])
for i in range(0,len(data_object.columns)):
    uni(data_object.columns[i])
           

根據之前的分析，No phone service，No internet service 可以等同于No

#No phone service 轉換為 No
data['MultipleLines'] = data['MultipleLines'].replace('No phone service','No')
data['MultipleLines'].value_counts()
#No internet service 轉換為 No
loc = ['OnlineSecurity', 'OnlineBackup', 'DeviceProtection', 'TechSupport', 'StreamingTV', 'StreamingMovies']
for i in loc:
    data[i] = data[i].replace('No internet service','No')
           

對隻有兩類值的特征标簽，進行二值化處理，用1， 0來代替

# Male= 1   Female = 0
data['gender'] = data['gender'].replace({'Male':1,'Female':0})

# Yes = 1   No = 0
loc = ['Partner','Dependents','PhoneService','MultipleLines','OnlineSecurity','OnlineBackup','DeviceProtection','TechSupport','StreamingTV','StreamingMovies','PaperlessBilling','Churn']
for i in loc:
    data[i] = data[i].replace({'Yes':1,'No':0})
           

數值型資料标準化處理

data[['tenure']] = StandardScaler().fit_transform(data[['tenure']])
data[['MonthlyCharges']] = StandardScaler().fit_transform(data[['MonthlyCharges']])
data[['TotalCharges']] = StandardScaler().fit_transform(data[['TotalCharges']])
           

對 InternetService，Contract，PaymentMethod 這三列進行獨熱編碼處理

#轉碼
InternetService = pd.get_dummies(data['InternetService'])
Contract = pd.get_dummies(data['Contract'])
PaymentMethod = pd.get_dummies(data['PaymentMethod'])
#合并資料
train = pd.concat([data,InternetService,Contract,PaymentMethod],axis=1)
#删除原标簽列
train = train.drop(['InternetService','Contract','PaymentMethod'],axis=1)
           

根據相關性系數，删除對預測結果影響較小的特征值

# 由于資料集不大可使用 corr() 計算出沒對屬性之間的标準相關系數  也稱為“皮爾遜系數”

corr_matrix = train.corr()
abs(corr_matrix['Churn']).sort_values(ascending=False)
           

根據情況來取舍相關性低的标簽列

本次案例選擇删除小于0.1的标簽

#删除列
loc = ['gender','PhoneService','MultipleLines','StreamingMovies','StreamingTV','DeviceProtection','OnlineBackup','Mailed check']
train= train.drop(loc,axis=1)
           

提取目标标簽，劃分資料

train_x = train.drop(['Churn'],axis=1)
train_y = train['Churn']
           

模型預測

導入模型包，

from sklearn.linear_model import LogisticRegression  #邏輯回歸
from sklearn.ensemble import RandomForestClassifier  #随機森林
from xgboost import XGBClassifier #xgboost 
#導入評分子產品
from sklearn.metrics import recall_score  
from sklearn.metrics import precision_score
from sklearn.metrics import f1_score

           

訓練模型并選擇最優模型

# 定義一個函數，使其能輸入訓練資料和模型
#然後傳回模型分數
def model_fit(x,y,model):
    model = model
    model.fit(x,y)
    predict = model.predict(x)

    recall = recall_score(y,predict)
    precision = precision_score(y,predict)
    f1 = f1_score(y,predict)

    print(model)
    print('recall   :{:.3f}'.format(recall))
    print('precision:{:.3f}'.format(precision))
    print('f-1      :{:.3f}'.format(f1))
           

訓練模型

LR = LogisticRegression()
RF = RandomForestClassifier()
XG = XGBClassifier(eval_metric=['logloss','auc','error'],objective ='reg:squarederror')

model_fit(train_x,train_y,LR)
model_fit(train_x,train_y,RF)
model_fit(train_x,train_y,XG)
           

模型評估

LogisticRegression
recall   :0.547
precision:0.658
f-1      :0.597
RandomForestClassifier
recall   :0.994
precision:0.995
f-1      :0.994
XGBClassifier
recall   :0.819
precision:0.935
f-1      :0.873
           

根據精确率，召回率，F1值 綜合比較選擇最優模型

本次測試選擇模型為RandomForestClassifier

使用網格搜尋尋找最優參數

from sklearn.model_selection import GridSearchCV

param_grid = [
    {'n_estimators':[3,5,10,15,25,30],'max_features':[2,4,6,8]},
    {'bootstrap':[False],'n_estimators':[3,10],'max_features':[2,3,4]}
]



grid_search = GridSearchCV(RF,
                           param_grid,cv=5,
                           scoring='neg_mean_squared_error',
                           return_train_score=True)
           

grid_search.fit(train_x,train_y)
#列印最優參數
grid_search.best_params_
           

預測結果展示

final_model = grid_search.best_estimator_
final_predictions = final_model.predict_proba(train_x[:20])

output = pd.DataFrame({'customerID':data_id[:20],'ratio':final_predictions[:,1]})
output = output.sort_values('ratio',ascending=False)
output
           

建議：預測值越大的使用者流失率就越高，對于公司來說，可以根據實際情況設定流失率門檻值，達到門檻值的群體進行挽留措施，值越大優先級越高。

結果僅供參考

由于資料有限，測試和訓練使用的都是一樣的數，是以結果的過拟合問題較嚴重，方法思路是差不多的。