python， 在信用评级中，计算KS statistic值

# -*- coding: utf-8 -*-

import pandas as pd

from sklearn.grid_search import GridSearchCV

from sklearn.model_selection import train_test_split

from sklearn.preprocessing import StandardScaler

from sklearn.utils import shuffle

import numpy as np

from sklearn import metrics

from sklearn.metrics import log_loss, recall_score, precision_score, accuracy_score,f1_score

from sklearn.metrics import roc_curve, precision_recall_curve, roc_auc_score

# from sklearn.model_selection import cross_val_score

import lightgbm

def ks_statistic(Y,Y_hat):

    data = {"Y":Y,"Y_hat":Y_hat}

    df = pd.DataFrame(data)

    bins = np.array([-0.1,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0])

    category = pd.cut(df["Y_hat"],bins=bins)

    category = category.sort_values()

    #max_index = len(np.unique(df["Y_hat"]))

    Y = df.ix[category.index,:]['Y']

    Y_hat = df.ix[category.index,:]['Y_hat']

    df2 = pd.concat([Y,Y_hat],axis=1)

    df3 = pd.pivot_table(df2,values = ['Y_hat'],index ='Y_hat',columns='Y',aggfunc=len,fill_value=0)

    df4 = np.cumsum(df3)

    df5 = df4/df4.iloc[:,1].max()

    ks = max(abs(df5.iloc[:,0] - df5.iloc[:,1]))

    return ks/len(bins)

df = pd.read_csv('DC_ALL_20170217.csv', header=0)

X = df[df.columns.drop(['user_id','overdue'])].fillna(-999)

# X = df[['count','time_stamp','credit_limit','credit_card_use_rate','credit_count_x','bank_count','sex','occupation','education','marriage','hukou']]

y = df['overdue']

train = X.head(55596)

test = X.tail(69495-55596)

train_label = y.head(55596).convert_objects(convert_numeric=True)

X_train, X_test, y_train, y_test = train_test_split(\

	 train.values,  train_label, test_size=0.2, random_state=42)

max_depth = 5

subsample=0.8

learning_rate=0.01

n_estimators=400

random_state=3

nthread=4

is_unbalance=True

objective ='binary'

LGBM = lightgbm.LGBMClassifier(max_depth=max_depth, learning_rate=learning_rate,

n_estimators=n_estimators, objective=objective,is_unbalance=is_unbalance, nthread=nthread,subsample=subsample)

LGBM.fit(X_train, y_train)

y_test_v = LGBM.predict(X_test)

y_test_p = LGBM.predict_proba(X_test)[:, 1]

print 'auc: ', roc_auc_score(y_test, y_test_p)

print 'log_loss: ', log_loss(y_test, y_test_p)

print 'precision: ', precision_score(y_test, y_test_v)

print 'recall: ', recall_score(y_test, y_test_v)

print 'accuracy: ', accuracy_score(y_test, y_test_v)

print 'f1_score: ', f1_score(y_test, y_test_v)

print 'ks_statistic: ', ks_statistic(y_test.values, y_test_v)