LAPRAS

Lapras is designed to make the model developing job easily and conveniently. It contains these functions below in one key operation: data exploratory analysis, feature selection, feature binning, data visualization, scorecard modeling(a logistic regression model with excellent interpretability), performance measure.

Let's get started.

Usage

1.Exploratory Data Analysis lapras.detect() lapras.quality() lapras.IV() lapras.VIF() lapras.PSI()

2.Feature Selection lapras.select() lapras.stepwise()

3.Binning lapras.Combiner() lapras.WOETransformer() lapras.bin_stats() lapras.bin_plot()

4.Modeling lapras.ScoreCard()

5.Performance Measure lapras.perform() lapras.LIFT() lapras.score_plot() lapras.KS_bucket() lapras.PPSI() lapras.KS() lapras.AUC()

6.One Key Auto Modeling Lapras also provides a function which runs all the steps above automatically: lapras.auto_model()

Install

via pip

bash pip install lapras --upgrade -i https://pypi.org/simple

via source code

bash python setup.py install

install_requires = [ 'numpy >= 1.18.4', 'pandas >= 0.25.1', 'scipy >= 1.3.2', 'scikit-learn =0.22.2', 'seaborn >= 0.10.1', 'statsmodels >= 0.13.1', 'tensorflow >= 2.2.0', 'hyperopt >= 0.2.7', 'pickle >= 4.0', ]

Documents

```python import lapras

import pandas as pd import numpy as np from sklearn.model_selection import train_test_split import matplotlib as mpl import matplotlib.pyplot as plt

pd.options.display.max_colwidth = 100 import math %matplotlib inline ```

```python

Read in data file

df = pd.read_csv('data/demo.csv',encoding="utf-8") ```

python to_drop = ['id'] # exclude the features which not being used， eg:id target = 'bad' # Y label name train_df, test_df, _, _ = train_test_split(df, df[[target]], test_size=0.3, random_state=42) # to divide the training set and testing set, strongly recommended

```python

EDA(Exploratory Data Analysis)

Parameter details：

dataframe=None

lapras.detect(train_df).sort_values("missing") ```

```python

Calculate IV value of features（Calculate by default decision tree binning）

Parameter details：

dataframe=None original data

target = 'target' Y label name

lapras.quality(train_df.drop(to_drop,axis=1),target = target) ```

```python

Calculate PSI betweet features

Parameter details：

actual=None actual feature

predict=None prediction feature

bins=10 count of binning

return_frame=False return the dataframe of binning if set to true

cols = list(lapras.quality(train_df,target = target).reset_index()['index']) for col in cols: if col not in [target]: print("%s: %.4f" % (col,lapras.PSI(train_df[col], test_df[col]))) python score: 0.1500 age: 0.0147 wealth: 0.0070 education: 0.0010 max_unpay_day: 0.0042 id: 0.0000 period: 0.0030 python

Calculate VIF

Parameter details：

dataframe=None

lapras.VIF(train_df.drop(['id','bad'],axis=1)) python wealth 1.124927 max_unpay_day 2.205619 score 18.266471 age 17.724547 period 1.193605 education 1.090158 dtype: float64 python

Calculate IV value

Parameter details：

feature=None feature data

target=None Y label data

lapras.IV(train_df['age'],train_df[target]) python 0.5045879202656338 python

Features filtering

Parameter details：

frame=None original data

target=None Y label name

empty=0.9 empty feature filtering， feature will be removed if data missing ratio greater than the threshold

iv=0.02 IV value filtering， feature will be removed if IV value lesser than the threshold

corr=0.7 correlation filtering， feature will be removed if correlation value greater than the threshold

vif=False multicollinearity filtering， feature will be removed if multicollinearity value greater than the threshold, default False due to a large number of calculations

return_drop=False reture the removed features if set to true

exclude=None features will be remained if set into this parameter

train_selected, dropped = lapras.select(train_df.drop(to_drop,axis=1),target = target, empty = 0.95, \ iv = 0.05, corr = 0.9, vif = False, return_drop=True, exclude=[]) print(dropped) print(train_selected.shape) train_selected python {'empty': array([], dtype=float64), 'iv': array([], dtype=object), 'corr': array([], dtype=object)} (5502, 7) ```

```python

Feature Binning， following methods are supported: monotonous binning， decision tree binning， Kmeans binning， equal frequency binning， equal step size binning

Parameter details：

X=None original data

y=None Y label name

method='dt' Binning method：'dt':decision tree binning(default),'mono':monotonous binning,'kmeans':Kmeans binning,'quantile':equal frequency binning,'step':equal step size binning

min_samples=1 the least sample numbers in each binning， represent the count of numbers when greater than 1， represent the ratio of total count when between 0 and 1

n_bins=10 maximun binning count

c.load(dict) adjust the binning by loading a customized dict

c.export() export current binning information by dict format

c = lapras.Combiner() c.fit(train_selected, y = target,method = 'mono', min_samples = 0.05,n_bins=8) #empty_separate = False

# c.load({'age': [22.5, 23.5, 24.5, 25.5, 28.5,36.5],

# 'education': [ 3.5],

# 'max_unpay_day': [59.5],

# 'period': [5.0, 9.0],

# 'score': [205.5, 236.5, 265.5, 275.5, 294.5, 329.5, 381.5],

# 'wealth': [2.5, 3.5, 6.5]})

c.export() python {'age': [23.0, 24.0, 25.0, 26.0, 28.0, 29.0, 37.0], 'education': [3.0, 4.0], 'max_unpay_day': [171.0], 'period': [6.0, 10.0], 'score': [237.0, 272.0, 288.0, 296.0, 330.0, 354.0, 384.0], 'wealth': [3.0, 4.0, 5.0, 7.0]} python

To transform the original data into binning data

Parameter details：

X=None original data

labels=False binning label will be shown when set to true

c.transform(train_selected, labels=True).iloc[0:10,:] ```

```python

To output bin_stats and bin_plot

Parameter details：

frame=None data transformed by Combiner, keeping binning labels

col=None features to be outputed

target='target' Y label name

Note：The binning details may be different between traning set and testing set due to Population Stability.

cols = list(lapras.quality(train_selected,target = target).reset_index()['index']) for col in cols: if col != target: print(lapras.bin_stats(c.transform(train_selected[[col, target]], labels=True), col=col, target=target)) lapras.bin_plot(c.transform(train_selected[[col,target]], labels=True), col=col, target=target) python score bad_count total_count bad_rate ratio woe \ 0 00.[-inf,237.0) 136 805 0.168944 0.146310 0.944734 1 01.[237.0,272.0) 101 832 0.121394 0.151218 0.558570 2 02.[272.0,288.0) 46 533 0.086304 0.096874 0.178240 3 03.[288.0,296.0) 20 295 0.067797 0.053617 -0.083176 4 04.[296.0,330.0) 73 1385 0.052708 0.251727 -0.350985 5 05.[330.0,354.0) 18 812 0.022167 0.147583 -1.248849 6 06.[354.0,384.0) 8 561 0.014260 0.101963 -1.698053 7 07.[384.0,inf) 1 279 0.003584 0.050709 -3.089758

 iv  total_iv

0 0.194867 0.735116 1 0.059912 0.735116 2 0.003322 0.735116 3 0.000358 0.735116 4 0.026732 0.735116 5 0.138687 0.735116 6 0.150450 0.735116 7 0.160788 0.735116 ![png](http://img.badtom.cn/output_13_1.png)python age bad_count total_count bad_rate ratio woe \ 0 00.[-inf,23.0) 90 497 0.181087 0.090331 1.028860 1 01.[23.0,24.0) 77 521 0.147793 0.094693 0.785844 2 02.[24.0,25.0) 57 602 0.094684 0.109415 0.280129 3 03.[25.0,26.0) 38 539 0.070501 0.097964 -0.041157 4 04.[26.0,28.0) 58 997 0.058175 0.181207 -0.246509 5 05.[28.0,29.0) 20 379 0.052770 0.068884 -0.349727 6 06.[29.0,37.0) 57 1657 0.034400 0.301163 -0.796844 7 07.[37.0,inf) 6 310 0.019355 0.056343 -1.387405

 iv  total_iv

0 0.147647 0.45579 1 0.081721 0.45579 2 0.009680 0.45579 3 0.000163 0.45579 4 0.009918 0.45579 5 0.007267 0.45579 6 0.137334 0.45579 7 0.062060 0.45579 ![png](http://img.badtom.cn/output_13_3.png)python wealth bad_count total_count bad_rate ratio woe \ 0 00.[-inf,3.0) 106 593 0.178752 0.107779 1.013038 1 01.[3.0,4.0) 84 1067 0.078725 0.193929 0.078071 2 02.[4.0,5.0) 88 1475 0.059661 0.268084 -0.219698 3 03.[5.0,7.0) 99 1733 0.057126 0.314976 -0.265803 4 04.[7.0,inf) 26 634 0.041009 0.115231 -0.614215

 iv  total_iv

0 0.169702 0.236205 1 0.001222 0.236205 2 0.011787 0.236205 3 0.019881 0.236205 4 0.033612 0.236205 ```

```python education bad_count total_count bad_rate ratio woe \ 0 00.[-inf,3.0) 225 2123 0.105982 0.385860 0.405408 1 01.[3.0,4.0) 61 648 0.094136 0.117775 0.273712 2 02.[4.0,inf) 117 2731 0.042841 0.496365 -0.568600

 iv  total_iv

0 0.075439 0.211775 1 0.009920 0.211775 2 0.126415 0.211775 ![png](http://img.badtom.cn/output_13_7.png)python max_unpay_day bad_count total_count bad_rate ratio woe \ 0 00.[-inf,171.0) 330 5098 0.064731 0.926572 -0.132726 1 01.[171.0,inf) 73 404 0.180693 0.073428 1.026204

 iv  total_iv

0 0.015426 0.134699 1 0.119272 0.134699 ![png](http://img.badtom.cn/output_13_9.png)python period bad_count total_count bad_rate ratio woe \ 0 00.[-inf,6.0) 52 1158 0.044905 0.210469 -0.519398 1 01.[6.0,10.0) 218 2871 0.075932 0.521810 0.038912 2 02.[10.0,inf) 133 1473 0.090292 0.267721 0.227787

 iv  total_iv

0 0.045641 0.061758 1 0.000803 0.061758 2 0.015314 0.061758 ```

```python

WOE value transformation

transer.fit()：

X=None data transformed by Combiner

y=None Y label

exclude=None features exclude from transformation

transer.transform()：

X=None

transer.export()：

Note： Only training set need to be fit

transfer = lapras.WOETransformer() transfer.fit(c.transform(train_selected), train_selected[target], exclude=[target])

train_woe = transfer.transform(c.transform(train_selected)) transfer.export() python {'age': {0: 1.0288596439961428, 1: 0.7858440185299318, 2: 0.2801286322797789, 3: -0.041156782250006324, 4: -0.24650930955337075, 5: -0.34972695582581514, 6: -0.7968444812848496, 7: -1.387405073069694}, 'education': {0: 0.4054075821430197, 1: 0.27371220345368763, 2: -0.5685998002779383}, 'max_unpay_day': {0: -0.13272639517618706, 1: 1.026204224879801}, 'period': {0: -0.51939830439238, 1: 0.0389118677598222, 2: 0.22778739438526965}, 'score': {0: 0.9447339847162963, 1: 0.5585702161999536, 2: 0.17824043251497793, 3: -0.08317566500410743, 4: -0.3509853692471706, 5: -1.2488485442424984, 6: -1.6980533007340262, 7: -3.089757954582164}, 'wealth': {0: 1.01303813013795, 1: 0.0780708378046198, 2: -0.21969844672815222, 3: -0.2658032661768855, 4: -0.6142151848362123}} python

Features filtering could be done once more after transformed into WOE value. This is optional.

train_woe, dropped = lapras.select(train_woe,target = target, empty = 0.9, \ iv = 0.02, corr = 0.9, vif = False, return_drop=True, exclude=[]) print(dropped) print(train_woe.shape) train_woe.head(10) python {'empty': array([], dtype=float64), 'iv': array([], dtype=object), 'corr': array([], dtype=object)} (5502, 7) ```

```python

stepwise regression, to select best features, this is optional

Parameter details：

frame=None original data

target='target' Y label name

estimator='ols' model for regression, supporting 'ols', 'lr', 'lasso', 'ridge'

direction='both' direction for stepwise, supporting 'forward', 'backward', 'both'

criterion='aic' metric, supporting 'aic', 'bic', 'ks', 'auc'

max_iter=None max iteration times

return_drop=False return cols being removed if set to true

exclude=None exclude features

final_data = lapras.stepwise(train_woe,target = target, estimator='ols', direction = 'both', criterion = 'aic', exclude = []) final_data ```

```python

Scorecard modeling

Parameter details：

base_odds=1/60,base_score=600 When base_odds is 1/60, the corresponding base_score will be 600

pdo=40,rate=2 If the base_odds decrease by half, the corresponding pdo will increase by 40, these are the default parameters

combiner=None Combiner, input the fitted object

transfer=None WOETransformer, input the fitted object

ScoreCard.fit()：

X=None WOE value

y=None Y label

card = lapras.ScoreCard( combiner = c, transfer = transfer ) col = list(final_data.drop([target],axis=1).columns) card.fit(final_data[col], final_data[target])

python ScoreCard(base_odds=0.016666666666666666, base_score=600, card=None, combiner=, pdo=40, rate=2, transfer=) python

ScoreCard class method expaination

ScoreCard.predict() predict score for each sample：

X=None

ScoreCard.predict_prob() predict prob for each sample：

X=None

ScoreCard.export() output the details of scorecard by dict format

ScoreCard.get_params() to get the parameters of scorecard by dict format, usually used in deployment

card.intercept_ intercept of logical regression

card.coef_ coefficient of logical regression

final_result = final_data[[target]].copy() score = card.predict(final_data[col]) prob = card.predict_prob(final_data[col])

final_result['score'] = score final_result['prob'] = prob print("card.intercept_:%s" % (card.intercept_)) print("card.coef_:%s" % (card.coef_)) card.get_params()['combiner'] card.get_params()['transfer'] card.export() python card.intercept_:-2.5207582925622476 card.coef_:[0.32080944 0.3452988 0.68294643 0.66842902]

{'age': {'[-inf,23.0)': -39.69, '[23.0,24.0)': -30.31, '[24.0,25.0)': -10.81, '[25.0,26.0)': 1.59, '[26.0,28.0)': 9.51, '[28.0,29.0)': 13.49, '[29.0,37.0)': 30.74, '[37.0,inf)': 53.52}, 'intercept': {'[-inf,inf)': 509.19}, 'max_unpay_day': {'[-inf,171.0)': 2.64, '[171.0,inf)': -20.45}, 'score': {'[-inf,237.0)': -37.23, '[237.0,272.0)': -22.01, '[272.0,288.0)': -7.02, '[288.0,296.0)': 3.28, '[296.0,330.0)': 13.83, '[330.0,354.0)': 49.22, '[354.0,384.0)': 66.92, '[384.0,inf)': 121.77}, 'wealth': {'[-inf,3.0)': -18.75, '[3.0,4.0)': -1.45, '[4.0,5.0)': 4.07, '[5.0,7.0)': 4.92, '[7.0,inf)': 11.37}} python

model performance metrics, including KS, AUC, ROC curve, KS curve, PR curve

Parameter details

feature=None predicted value

target=None actual label

lapras.perform(prob,final_result[target]) python KS: 0.4160 AUC: 0.7602 ```

```python

Parameter details

frame=None original dataframe

score='score' score label name

target='target' Y label name

score_bond=None score boundary, default by 30, customized by list, e.g. [100,200,300]

lapras.score_plot(final_result,score='score', target=target) python bad: [42, 78, 70, 104, 61, 28, 18, 1, 1, 0] good: [129, 249, 494, 795, 1075, 972, 825, 282, 164, 114] all: [171, 327, 564, 899, 1136, 1000, 843, 283, 165, 114] all_rate: ['3.11%', '5.94%', '10.25%', '16.34%', '20.65%', '18.18%', '15.32%', '5.14%', '3.00%', '2.07%'] bad_rate: ['24.56%', '23.85%', '12.41%', '11.57%', '5.37%', '2.80%', '2.14%', '0.35%', '0.61%', '0.00%'] ```

```python

LIFT show

feature=None predicted value

target=None actual label

recall_list=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1] default

lapras.LIFT(prob,final_data[target]) ```

Automatical modeling

```python

auto_model parameters df,target,to_drop are necessary, others are optional

bins_show=False showing the binning graphs when set to true

iv_rank=False feature IV values will be ranked when set to true

perform_show=False showing performance(training set)

coef_negative=True coefficient can be negative if set to true

return: ScoreCard object

auto_card = lapras.auto_model(df=train_df,target=target,to_drop=to_drop,bins_show=False,iv_rank=False,perform_show=False, coef_negative = False, empty = 0.95, iv = 0.02, corr = 0.9, vif = False, method = 'mono', n_bins=8, min_samples=0.05, pdo=40, rate=2, base_odds=1 / 60, base_score=600) python ——data filtering—— original feature：6 filtered features：6

——feature binning——

——WOE value transformation——

——feature filtering once more—— original feature：6 filtered features：6

——scorecard modeling—— intercept: -2.520670026708529 coef: [0.66928671 0.59743968 0.31723278 0.22972838 0.28750881 0.26435224]

——model performance metrics—— KS: 0.4208 AUC: 0.7626 recall precision improve 0 0.1 0.238095 3.188586 1 0.2 0.254777 3.411990 2 0.3 0.239521 3.207679 3 0.4 0.193742 2.594611 4 0.5 0.182805 2.448141 5 0.6 0.171510 2.296866 6 0.7 0.160501 2.149437 7 0.8 0.130259 1.744435 8 0.9 0.110603 1.481206 9 1.0 0.074671 1.000000

Automatic modeling finished, time costing： 0 second ```