skcredit 0.0.1

scorecard

scikit-credit

scikit-credit 是用于信贷风控评分卡建模的自动化工具。如下是 scikit-credit 的简介：

• 自动分箱
• 特征选择
• 模型训练
• 评分生成

特性

• 自动分箱：分箱算法使用以信息量（IV）为分裂标准的分类与回归树（CART）。分类与回归树（CART）分裂前自动探测特征与特征WOE之间的单调关系，分裂过程中借鉴了XGBoost限制单调性的实现，使得分箱结果特征与特征WOE单调。
• 特征选择：通过分箱结果的检查与信息论为标准的特征选择算法进行特征选择。
• 模型训练：使用以KS为标准的逐步回归，保证入模特征回归系数为正且统计显著。

玩家

信贷风控模型分析师

示例

完整代码请参见 example.ipynb

单特征分箱

分类单特征分箱

默认参数

from skcredit.feature_discretization import SplitCat

# min_bin_cnt_negative=75
# min_bin_cnt_positive=75
# min_information_value_split_gain=0.015

sc = SplitCat()
sc.fit(dataset["EDUCATION"], dataset["default.payment.next.month"])
print(sc.table.to_markdown())

	Column	Bucket	CntPositive	CntNegative	WoE	IvS
0	EDUCATION	{0, 1, 4, 5}	1529	6756	-0.229193	0.0181077
1	EDUCATION	{2, 3, 6}	3456	10759	0.120993	0.00955926
2	EDUCATION	{nan}	0	0	0	0

调整参数

from skcredit.feature_discretization import SplitNum

sc = SplitCat(min_information_value_split_gain=0.0001)
sc.fit(dataset["EDUCATION"], dataset["default.payment.next.month"])
print(sc.table.to_markdown())

	Column	Bucket	CntPositive	CntNegative	WoE	IvS
0	EDUCATION	{0, 1, 4, 5, 6}	2069	8984	-0.209693	0.0152528
1	EDUCATION	{2}	3330	10700	0.0914156	0.00400755
2	EDUCATION	{3}	1237	3680	0.168463	0.00486862
3	EDUCATION	{nan}	0	0	0	0

连续单特征分箱

默认参数

from skcredit.feature_discretization import SplitNum

# min_bin_cnt_negative=75
# min_bin_cnt_positive=75
# min_information_value_split_gain=0.015

sn = SplitNum()
sn.fit(dataset["LIMIT_BAL"], dataset["default.payment.next.month"])
print(sn.table.to_markdown())

	Column	Bucket	CntPositive	CntNegative	WoE	IvS
0	LIMIT_BAL	(-inf,40000.0]	1555	2756	0.686382	0.0798734
1	LIMIT_BAL	(40000.0,140000.0]	2767	8212	0.170854	0.0111888
2	LIMIT_BAL	(140000.0,+inf)	2314	12396	-0.419709	0.0763266
3	LIMIT_BAL	[nan,nan]	0	0	0	0

调整参数

from skcredit.feature_discretization import SplitNum

sn = SplitNum(min_information_value_split_gain=0.0001)
sn.fit(dataset["LIMIT_BAL"], dataset["default.payment.next.month"])
print(sn.table.to_markdown())

	Column	Bucket	CntPositive	CntNegative	WoE	IvS
0	LIMIT_BAL	(-inf,10000.0]	197	296	0.851531	0.0144909
1	LIMIT_BAL	(10000.0,40000.0]	1358	2460	0.664539	0.0660227
2	LIMIT_BAL	(40000.0,70000.0]	1328	3593	0.263374	0.0122039
3	LIMIT_BAL	(70000.0,120000.0]	1112	3468	0.121269	0.00232077
4	LIMIT_BAL	(120000.0,140000.0]	327	1151	0.000260766	3.35032e-09
5	LIMIT_BAL	(140000.0,220000.0]	1103	5184	-0.288856	0.0160792
6	LIMIT_BAL	(220000.0,240000.0]	223	1133	-0.366765	0.00546071
7	LIMIT_BAL	(240000.0,260000.0]	138	733	-0.411205	0.00434948
8	LIMIT_BAL	(260000.0,360000.0]	556	3164	-0.480137	0.0247926
9	LIMIT_BAL	(360000.0,460000.0]	165	1160	-0.691543	0.0171397
10	LIMIT_BAL	(460000.0,+inf)	129	1022	-0.811017	0.0197102
11	LIMIT_BAL	[nan,nan]	0	0	0	0

多特征分箱

手动分箱

from skcredit.feature_discretization import DiscreteCust

sc = SplitCat(min_information_value_split_gain=0.0001)
sc.fit(train_x["EDUCATION"], train_y)
sn = SplitNum(min_information_value_split_gain=0.0001)
sn.fit(train_x["LIMIT_BAL"], train_y)

cust = DiscreteCust(keep_columns=["ID"], date_columns=[], cat_spliter={"EDUCATION": sc}, num_spliter={"LIMIT_BAL":sn})
cust.fit(train_x, train_y)

cust = DiscreteCust(keep_columns=["ID"], date_columns=[], cat_spliter={"EDUCATION": sc}, num_spliter={"LIMIT_BAL":sn})
cust.fit(train_x, train_y)

自动分箱

auto = DiscreteAuto(keep_columns=["ID"], date_columns=[], cat_columns=cat_columns, num_columns=num_columns)
auto.fit(train_x, train_y)
train_x = auto.transform(train_x)
test_x  = auto.transform(test_x )

查看分箱后信息

# cust.information_value_score.head() 
print(auto.information_value_score.head().to_markdown())

	IvS
PAY_0	0.874626
PAY_2	0.560568
PAY_3	0.420644
PAY_4	0.364907
PAY_5	0.330149

# cust.information_value_table.head()
print(auto.information_value_table.head().to_markdown())

	Column	Bucket	CntPositive	CntNegative	WoE	IvS
0	PAY_0	(-inf,0]	2391	14987	-0.578847	0.217663
1	PAY_0	(0,1]	959	1821	0.615374	0.0544047
2	PAY_0	(1,+inf)	1635	707	2.09499	0.602558
3	PAY_0	[nan,nan]	0	0	0	0
0	PAY_2	(-inf,0]	3105	16066	-0.387067	0.113953

特征选择

select = SelectBins(keep_columns=["ID"], date_columns=[])
select.fit(train_x, train_y)
train_x = select.transform(train_x)
test_x  = select.transform(test_x )

select = SelectCIFE(keep_columns=["ID"], date_columns=[], nums_feature=10)
select.fit(train_x, train_y)
train_x = select.transform(train_x)
test_x  = select.transform(test_x )

模型训练

lmclassifier = LMClassifier(keep_columns=["ID"], date_columns=[])
lmclassifier.fit(train_x, train_y)
print("train ks {}".format(lmclassifier.score(train_x, train_y)))
print("test  ks {}".format(lmclassifier.score(test_x,  test_y )))

train ks 0.41442
test ks 0.3914

评分卡生成

lmcreditcard = LMCreditcard(
        keep_columns=["ID"], date_columns=[], discrete=discrete, lmclassifier=lmclassifier, BASE=500,  PDO=20,  ODDS=1)
lmcreditcard.show_scorecard()

	Column	Bucket	WoE	Coefficients	PartialScore	OffsetScores
0	PAY_AMT2	(-inf,30.0]	0.555291	0.284123	-4.55231	536.095
1	PAY_AMT2	(30.0,4931.0]	0.0159422	0.284123	-0.130696	536.095
2	PAY_AMT2	(4931.0,15282.0]	-0.424907	0.284123	3.48342	536.095
3	PAY_AMT2	(15282.0,+inf)	-1.21984	0.284123	10.0003	536.095
4	PAY_AMT2	[nan,nan]	0	0.284123	-0	536.095
0	PAY_AMT1	(-inf,10.0]	0.688936	0.283557	-5.63668	536.095
1	PAY_AMT1	(10.0,4861.0]	-0.0125914	0.283557	0.10302	536.095
2	PAY_AMT1	(4861.0,+inf)	-0.591584	0.283557	4.84017	536.095
3	PAY_AMT1	[nan,nan]	0	0.283557	-0	536.095
0	PAY_0	(-inf,0]	-0.578847	0.749677	12.5211	536.095
1	PAY_0	(0,1]	0.615374	0.749677	-13.3112	536.095
2	PAY_0	(1,+inf)	2.09499	0.749677	-45.317	536.095
3	PAY_0	[nan,nan]	0	0.749677	-0	536.095
0	PAY_3	(-inf,0]	-0.318517	0.204782	1.88204	536.095
1	PAY_3	(0,+inf)	1.36739	0.204782	-8.07961	536.095
2	PAY_3	[nan,nan]	0	0.204782	-0	536.095
0	PAY_2	(-inf,0]	-0.387067	0.0994362	1.11054	536.095
1	PAY_2	(0,+inf)	1.51702	0.0994362	-4.35252	536.095
2	PAY_2	[nan,nan]	0	0.0994362	-0	536.095
0	PAY_4	(-inf,0]	-0.265818	0.331433	2.54205	536.095
1	PAY_4	(0,+inf)	1.41463	0.331433	-13.5283	536.095
2	PAY_4	[nan,nan]	0	0.331433	-0	536.095
0	PAY_AMT4	(-inf,0.0]	0.475069	0.407116	-5.58058	536.095
1	PAY_AMT4	(0.0,1813.0]	0.0656848	0.407116	-0.771592	536.095
2	PAY_AMT4	(1813.0,4000.0]	-0.128344	0.407116	1.50764	536.095
3	PAY_AMT4	(4000.0,+inf)	-0.51447	0.407116	6.04342	536.095
4	PAY_AMT4	[nan,nan]	0	0.407116	-0	536.095