8.4. Pipelines

Pipelines are just a series of steps you perform on data in sklearn. (The sklearn guide to them is here.)

A “typical” pipeline in ML projects

  1. Preprocesses the data to clean and tranform variables

  2. Possibly selects a subset of variables from among the features to avoid overfitting (see also this)

  3. Runs a model on those cleaned variables

Tip

You can set up pipelines with make_pipeline.

8.4.1. Intro to pipes

For example, here is a simple pipeline:

from sklearn.pipeline import make_pipeline 
from sklearn.impute import SimpleImputer
from sklearn.linear_model import Ridge

# set_config(pandas) might slow down sklearn, but should be used during development to 
# facilitate EDA/ABCD, bc we can see the transformed data with variable names
from sklearn import set_config          
set_config(transform_output="pandas") 

ridge_pipe = make_pipeline(SimpleImputer(),Ridge(1.0))

You put a series of steps inside make_pipeline, separated by commas.

The pipeline object (printed out below) is a list of steps, where each step has a name (e.g. “simpleimputer” ) and a task associated with that name (e.g. “SimpleImputer()”).

ridge_pipe
Pipeline(steps=[('simpleimputer', SimpleImputer()), ('ridge', Ridge())])
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Tip

You can .fit() and .predict() pipelines like any model, and they can be used in cross_validate too!

Using it is the same as using any estimator! After I load the data we’ve been using from the last two pages below (hidden), we can fit and predict like on the “one model intro” page:

import pandas as pd
import numpy as np
from sklearn.linear_model import Ridge
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold, cross_validate

url        = 'https://github.com/LeDataSciFi/data/blob/main/Fannie%20Mae/Fannie_Mae_Plus_Data.gzip?raw=true'
fannie_mae = pd.read_csv(url,compression='gzip').dropna()
y          = fannie_mae.Original_Interest_Rate
fannie_mae = (fannie_mae
                  .assign(l_credscore = np.log(fannie_mae['Borrower_Credit_Score_at_Origination']),
                          l_LTV = np.log(fannie_mae['Original_LTV_(OLTV)']),
                         )
              .iloc[:,-11:] 
             )

rng = np.random.RandomState(0) # this helps us control the randomness so we can reproduce results exactly
X_train, X_test, y_train, y_test = train_test_split(fannie_mae, y, random_state=rng)
ridge_pipe.fit(X_train,y_train)
ridge_pipe.predict(X_test)
array([5.95256433, 4.20060942, 3.9205946 , ..., 4.06401663, 5.30024985,
       7.32600213])

Those are the same numbers as before - good!

We can use this pipeline in our cross validation in place of the estimator:

cross_validate(ridge_pipe,X_train,y_train,
              cv=KFold(5), scoring='r2')['test_score'].mean()
0.9030537085469961

8.4.2. Preprocessing in pipes

Warning

(Virtually) All preprocessing should be done in the pipeline!

This is the link you should start with to see how you might clean and preprocess data. Key preprocessing steps include

  • Filling in missing values (imputation) or dropping those observations

  • Standardization

  • Encoding categorical data

With real-world data, you’ll have many data types. So the preprocessing steps you apply to one column won’t necessarily be what the next column needs.

I use ColumnTransformer to assemble my preprocessing portion of my full pipeline, and it allows me to process different variables differently.


The generic steps to preprocess in a pipeline:

  1. Set up a pipeline for numerical data

  2. Set up a pipeline for categorical variables

  3. Set up the ColumnTransformer:

    • ColumnTransformer() is a function, so it needs the parentheses “()”

    • The first argument inside it is a list (so now it is ColumnTransformer([]))

    • Each element in that list is a tuple that has three parts:

      • name of the step (you decide the name),

      • estimator/pipeline to use on that step,

      • and which variables to use it on

    • Put the pipeline for each variable type as its own tuple inside ColumnTransformer([<here!>])

  4. Use the ColumnTransformer set as the first step inside your glorious estimation pipeline.


So, let me put this together:

Tip

This is good pseudo!

from sklearn.preprocessing import OneHotEncoder 
from sklearn.compose import ColumnTransformer, make_column_selector

#############
# Step 1: how to deal with numerical vars
# pro-tip: you might set up several numeric pipelines, because
# some variables might need very different treatment!
#############

numer_pipe = make_pipeline(SimpleImputer()) 
# this deals with missing values (somehow?)
# you might also standardize the vars in this numer_pipe

#############
# Step 2: how to deal with categorical vars
#############

cat_pipe   = make_pipeline(OneHotEncoder(drop='first',sparse_output=False))

# notes on this cat pipe:
#     OneHotEncoder is just one way to deal with categorical vars
#     drop='first' is necessary if the model is regression
#     sparse_output=False might slow down sklearn, BUT IT MUST BE USED WITH set_config(pandas) !!! 

#############
# Step 3: combine the subparts
#############

preproc_pipe = ColumnTransformer(  
    [ # arg 1 of ColumnTransformer is a list, so this starts the list
    # a tuple for the numerical vars: name, pipe, which vars to apply to
    ("num_impute", numer_pipe, ['l_credscore','TCMR']),
    # a tuple for the categorical vars: name, pipe, which vars to apply to
    ("cat_trans", cat_pipe, ['Property_state'])
    ]
    , remainder = 'drop' # you either drop or passthrough any vars not modified above
)

#############
# Step 4: put the preprocessing into an estimation pipeline
#############

new_ridge_pipe = make_pipeline(preproc_pipe,Ridge(1.0))

The data loaded above has no categorical variables, so I’m going to reload the data and keep new variables to illustrate what we can do:

  • 'TCMR','l_credscore' are numerical

  • 'Property_state' is categorical

  • 'l_LTV' is in the data, but should be dropped (because of remainder='drop')

So here is the raw data:

url        = 'https://github.com/LeDataSciFi/data/blob/main/Fannie%20Mae/Fannie_Mae_Plus_Data.gzip?raw=true'
fannie_mae = pd.read_csv(url,compression='gzip').dropna()
y          = fannie_mae.Original_Interest_Rate
fannie_mae = (fannie_mae
                  .assign(l_credscore = np.log(fannie_mae['Borrower_Credit_Score_at_Origination']),
                          l_LTV = np.log(fannie_mae['Original_LTV_(OLTV)']),
                         )
             [['TCMR', 'Property_state', 'l_credscore', 'l_LTV']]
             )

rng = np.random.RandomState(0) # this helps us control the randomness so we can reproduce results exactly
X_train, X_test, y_train, y_test = train_test_split(fannie_mae, y, random_state=rng)

display(X_train.head())
display(X_train.describe().T.round(2))
TCMR Property_state l_credscore l_LTV
4326 4.651500 IL 6.670766 4.499810
15833 4.084211 TN 6.652863 4.442651
66753 3.675000 MO 6.635947 4.442651
23440 3.998182 MO 6.548219 4.553877
4155 4.651500 CO 6.602588 4.442651
count mean std min 25% 50% 75% max
TCMR 7938.0 3.36 1.29 1.50 2.21 3.00 4.45 6.66
l_credscore 7938.0 6.60 0.07 6.27 6.55 6.61 6.66 6.72
l_LTV 7938.0 4.51 0.05 4.25 4.49 4.50 4.55 4.57

We could .fit() and .transform() using the preproc_pipe from step 3 (or just .fit_transform() to do it in one command) to see how it transforms the data.

transformed_Xtrain = preproc_pipe.fit_transform(X_train)
transformed_Xtrain
num_impute__l_credscore num_impute__TCMR cat_trans__Property_state_AL cat_trans__Property_state_AR cat_trans__Property_state_AZ cat_trans__Property_state_CA cat_trans__Property_state_CO cat_trans__Property_state_CT cat_trans__Property_state_DC cat_trans__Property_state_DE ... cat_trans__Property_state_SD cat_trans__Property_state_TN cat_trans__Property_state_TX cat_trans__Property_state_UT cat_trans__Property_state_VA cat_trans__Property_state_VT cat_trans__Property_state_WA cat_trans__Property_state_WI cat_trans__Property_state_WV cat_trans__Property_state_WY
4326 6.670766 4.651500 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
15833 6.652863 4.084211 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
66753 6.635947 3.675000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
23440 6.548219 3.998182 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
4155 6.602588 4.651500 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
123118 6.650279 1.556522 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
69842 6.647688 2.416364 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
51872 6.507278 6.054000 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
128800 6.618739 2.303636 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
46240 6.639876 4.971304 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

7938 rows × 53 columns

Notice

  • The l_LTV column is gone!

  • The property state variable is now 50+ variables (one dummy for each state, and a few territories)

  • The numerical variables aren’t changed (there are no missing variables, so the imputation does nothing)

display(transformed_Xtrain
        .describe().T.round(2)
        .iloc[:7,:]) # only show a few variables for space...
count mean std min 25% 50% 75% max
num_impute__l_credscore 7938.0 6.60 0.07 6.27 6.55 6.61 6.66 6.72
num_impute__TCMR 7938.0 3.36 1.29 1.50 2.21 3.00 4.45 6.66
cat_trans__Property_state_AL 7938.0 0.02 0.12 0.00 0.00 0.00 0.00 1.00
cat_trans__Property_state_AR 7938.0 0.01 0.10 0.00 0.00 0.00 0.00 1.00
cat_trans__Property_state_AZ 7938.0 0.03 0.17 0.00 0.00 0.00 0.00 1.00
cat_trans__Property_state_CA 7938.0 0.07 0.25 0.00 0.00 0.00 0.00 1.00
cat_trans__Property_state_CO 7938.0 0.03 0.16 0.00 0.00 0.00 0.00 1.00

8.4.3. Working with pipes

  • Using pipes is the same as any model: .fit() and .predict(), put into CVs

  • When modeling, you should spend time interrogating model predictions, plotting, and printing. Does the model struggle to predict certain observations? Does it excel at some?

  • You’ll want to tweak parts of your pipeline. The next pages cover how we can do that.