autoBinning 0.1.5.1

A small package for feature autoBinning

auto binning 分箱工具

安装

pip install autoBinning

基础工具 (simpleMethods)

my_list = [1,1,2,2,2,2,3,3,4,5,6,7,8,9,10,10,20,20,20,20,30,30,40,50,60,70,80,90,100]
my_list_y = [1,1,2,2,2,2,1,1,1,2,2,2,1,1]
t = simpleMethods(my_list)
t.equalSize(3)
# 每个分箱样本数平均
print(t.bins) # [  1.           5.33333333  20.         100.        ]
# 等间距划分分箱
t.equalValue(4)
print(t.bins) # [  1.    25.75  50.5   75.25 100.  ]
# 基于numpy histogram分箱
t.equalHist(4)
print(t.bins) # [  1.    25.75  50.5   75.25 100.  ]

基于标签的有监督自动分箱

向前迭代方法 (forward method)

# load data
df = pd.read_csv('credit_old.csv')
df = df[['Age','target']]
df = df.dropna()

基于最大woe分裂分箱

在得到尽可能细粒度的细分箱之后，寻找上下分箱woe差异最大的初始切割点，并得到woe趋势，之后迭代找到下一个woe差异最大且趋势相同的切割点，直到满足woe差异不大于一个阈值或分箱数（切割点数）满足要求

t = forwardSplit(df['Age'], df['target'])
t.fit(sby='woe',minv=0.01,init_split=20)
print(t.bins) # [16. 25. 29. 33. 36. 38. 40. 42. 44. 46. 48. 50. 52. 54. 55. 58. 60. 63. 72. 94.]
t = forwardSplit(df['Age'], df['target'])
t.fit(sby='woe',num_split=4,init_split=20)
print(t.bins) # [16. 42. 44. 48. 50. 94.]
print("bin\twoe")
for i in range(len(t.bins)-1):
    v = t.value[(t.x < t.bins[i+1]) & (t.x >= t.bins[i])]
    woe = t._cal_woe(v)
    print((t.bins[i], t.bins[i+1]),woe)

bin	woe
(16.0, 25.0) 0.11373232830301286
(25.0, 42.0) 0.07217546872710079
(42.0, 50.0) 0.04972042405868509
(50.0, 72.0) -0.07172614369435065
(72.0, 94.0) -0.13778318584223453

基于最大iv分裂分箱

与最大woe分裂分箱方法类似，在得到尽可能细粒度的细分箱之后，寻找iv值最大的切割点，并得到woe趋势，之后迭代找到下一个iv最大且woe趋势相同的切割点，直到分箱数（切割点数）满足要求

# sby='woeiv'时考虑woe趋势，sby='iv'时不考虑woe趋势
t = forwardSplit(df['Age'], df['target'])
t.fit(sby='iv',minv=0.1,init_split=20)
print(t.bins) # [16. 25. 29. 33. 36. 38. 40. 42. 44. 46. 48. 50. 58. 60. 63. 94.]
t = forwardSplit(df['Age'], df['target'])
t.fit(sby='iv',num_split=4,init_split=20)
print(t.bins) # [16. 25. 33. 36. 38. 94.]
t.fit(sby='woeiv',num_split=4,init_split=20)
print(t.bins) # [16. 25. 33. 36. 38. 94.]

print("bin\twoe")
for i in range(len(t.bins)-1):
    v = t.value[(t.x < t.bins[i+1]) & (t.x >= t.bins[i])]
    woe = t._cal_woe(v)
    print((t.bins[i], t.bins[i+1]),woe)

bin	woe
(16.0, 25.0) 0.11373232830301286
(25.0, 33.0) 0.06679187564362839
(33.0, 40.0) 0.06638021747875023
(40.0, 50.0) 0.05894173616389541
(50.0, 94.0) -0.07934608583946329

向后迭代方法 (backward method)

基于最大iv合并分箱

迭代每次删除一个分箱切点，是去掉后整体iv最大

t = backwardSplit(df['Age'], df['target'])
t.fit(sby='iv',num_split=5)
print(t.bins) # [16.  17.5 18.5 85.5 95. ]

基于卡方检验的合并分箱

1. 得到尽可能细粒度的细分箱切点

2. 每个切点计算上下相邻分箱的卡方检验值

3. 将卡方检验值最低的两个分箱合并

4. 重复前两步直到达到分裂最小分裂切点数

1. First the input range is initialized by splitting it into sub-intervals with each sample getting own interval.

2. For every pair of adjacent sub-intervals a chi-square value is computed.

3. Merge pair with lowest chi-square into single bin.

4. Repeat 1 and 2 until number of bins meets predefined threshold.

t = backwardSplit(df['Age'], df['target'])
t.fit(sby='chi',num_split=7)
print(t.bins) # [16.  72.5 73.5 87.5 89.5 90.5 95. ]

基于spearman相关性做向后等频分箱

t = backwardSplit(df['Age'], df['target'])
t.fit_by_spearman(min_v=5, init_split=20)