外部数据评价函数

评价函数

此函数用于评价外部数据：主要指标有:PSI，相关系数、互信息指标、KS、GINI、IV、AUC、f1_score,此外还有其他描述性指标。

# psi score cal 
def psi_cal(score_col,label_col,date_col,data):
    '''
    score_col：str，外部数据列名称
    label_col：str，标签数据列名称
    date_col：str，日期数据名称
    data：dataFrame，df文件
    '''
    base_data = data[data[date_col] < 20210800]
    pre_data = data[data[date_col] >= 20210800]
    base_data_cut,bins = pd.qcut(base_data[score_col],10,duplicates='drop',retbins=True)# cut and return bins 
    base_bad_rate = base_data.flag.groupby(base_data_cut).mean() # cal base_data bin bad rate 
    pre_data_cut = pd.cut(pre_data[score_col],bins,include_lowest=True)
    pre_bad_rate = pre_data.flag.groupby(pre_data_cut).mean()  # cal pre_data bin bad rate 
    PSI = (base_bad_rate.values - pre_bad_rate.values)*np.log(base_bad_rate.values/pre_bad_rate.values)
    psi_value = PSI.sum()
    return psi_value

# 评价函数
def socre_test(test,label):
    '''
    单项数据的验证
    col：待检测数据 -- pd.Series
    label:Y值 -- pd.Series
    return : 各项描述指标和评价指标 -- dict
    '''
    # 计算查得率
    rate_response = (test.notna().sum()/len(test))*100  
    # 剔除未查到样本 
    test = test[test.notna()]
    label = label[test.index]
    # 描述性指标计算
    range_,min_,max_,mean_,std_,cov = test.max() -test.min(),test.min(),test.max(),test.mean(),test.std(),test.std()/test.mean()  # 描述统计指标
    # 相关系数指标 
    corr = test.corr(label)  # 
    # 互信息指标 
    mic_value= mic(test.values.reshape(-1,1),label.values.reshape(-1,1))
    mic_value = mic_value[0]
    # 按分位数进行分箱
    distribute_table = label.groupby(pd.qcut(test,10,duplicates='drop')).agg(['count','mean'])  
    # IV 计算 
    distribute_table.rename(columns={'count':'all_num','mean':'bad_rate'},inplace=True)
    distribute_table['good_rate'] = 1 - distribute_table['bad_rate']
    distribute_table['bad_num'] = distribute_table.all_num * distribute_table.bad_rate
    distribute_table['good_num'] = distribute_table.all_num * distribute_table.good_rate
    woe = np.log((distribute_table.bad_num/distribute_table.bad_num.sum())/(distribute_table.good_num/distribute_table.good_num.sum()))
    iv = ((distribute_table.bad_num/distribute_table.bad_num.sum()) - (distribute_table.good_num/distribute_table.good_num.sum()))*woe
    iv_value = iv.sum()

    # KS计算 
    ks_bad_rate = distribute_table['all_num']*distribute_table['bad_rate']/label.sum()
    ks_good_rate = (distribute_table['all_num']*distribute_table['good_rate'])/(len(label) - label.sum())
    ks_value = abs(ks_bad_rate.cumsum() - ks_good_rate.cumsum()).max()

    # Gini 系数 
    distribute_table['gini'] = 1 - ((distribute_table['bad_rate']*distribute_table['bad_rate']) + (distribute_table['good_rate']*distribute_table['good_rate']))
    gini_value = sum(distribute_table['gini']*distribute_table['all_num']/len(label))

    # AUC计算
    score = test.apply(lambda x:( x- min_)/(max_ - min_))
    if corr < 0:
        score = 1- score 
    
    auc_score = roc_auc_score(label[score.sort_values().index],score.sort_values())

    # f1 socre 
    if corr > 0:
        point = test.quantile(0.9)
        y_pred = test.apply(lambda x: 1 if x >= point else 0)
    else:
        point = test.quantile(0.1)
        y_pred = test.apply(lambda x: 1 if x <= point else 0)  
    f1score = f1_score(label,y_pred)
    return {'查得率':rate_response,'corr':corr,'mic':mic_value,'ks':ks_value,"iv":iv_value,"gini":gini_value,"auc":auc_score,'f1_socre':f1score,'极差':range_,'最大值':max_,"最小值":min_,"均值":mean_,"变异系数":cov }

}