Pipelines

Pipelines are just 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`.

Intro to pipes

{margin}
<img src="https://media.giphy.com/media/k5b6fkFnSA3yo/source.gif" alt="Mario" style="width:200px;">

For example, here is a simple pipeline:

In [1]:
from sklearn.pipeline import make_pipeline 
from sklearn.impute import SimpleImputer
from sklearn.linear_model import Ridge

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()").

In [2]:
ridge_pipe
Out[2]:
Pipeline(steps=[('simpleimputer', SimpleImputer()), ('ridge', Ridge())])
{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:

In [3]:
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/ledatascifi-2021/blob/main/data/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)
In [4]:
ridge_pipe.fit(X_train,y_train)
ridge_pipe.predict(X_test)
Out[4]:
array([5.95256433, 4.20060942, 3.9205946 , ..., 4.06401663, 5.30024985,
       7.32600213])

Those are the same numbers as before - good!

And we can use this pipeline in cross-validation method just like before:

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

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 functions, needs "()"
    • First argument 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!
In [6]:
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'))

# notes on this cat pipe:
#     OneHotEncoder is just one way to deal with categorical vars
#     drop='first' is necessary if the model is regression

#############
# 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 vars, so I'm going to reload the data and keep different variables:

  • '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:

In [7]:
url        = 'https://github.com/LeDataSciFi/ledatascifi-2021/blob/main/data/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. But the output is tough to use:

In [8]:
preproc_pipe.fit_transform(X_train)
Out[8]:
<7938x53 sparse matrix of type '<class 'numpy.float64'>'
	with 23792 stored elements in Compressed Sparse Row format>

So I added a convenience function (df_after_transform) to the community codebook to show the dataframe after the ColumnTransformer step.

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)

This is the transformed data:

In [9]:
from df_after_transform import df_after_transform

df_after_transform(preproc_pipe,X_train)
Out[9]:
l_credscore TCMR Property_state_AL Property_state_AR Property_state_AZ Property_state_CA Property_state_CO Property_state_CT Property_state_DC Property_state_DE ... Property_state_SD Property_state_TN Property_state_TX Property_state_UT Property_state_VA Property_state_VT Property_state_WA Property_state_WI Property_state_WV Property_state_WY
0 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
1 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
2 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
3 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
4 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
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
7933 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
7934 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
7935 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
7936 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
7937 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

In [10]:
display(df_after_transform(preproc_pipe,X_train)
        .describe().T.round(2)
        .iloc[:7,:]) # only show a few variables for space...
count mean std min 25% 50% 75% max
l_credscore 7938.0 6.60 0.07 6.27 6.55 6.61 6.66 6.72
TCMR 7938.0 3.36 1.29 1.50 2.21 3.00 4.45 6.66
Property_state_AL 7938.0 0.02 0.12 0.00 0.00 0.00 0.00 1.00
Property_state_AR 7938.0 0.01 0.10 0.00 0.00 0.00 0.00 1.00
Property_state_AZ 7938.0 0.03 0.17 0.00 0.00 0.00 0.00 1.00
Property_state_CA 7938.0 0.07 0.25 0.00 0.00 0.00 0.00 1.00
Property_state_CO 7938.0 0.03 0.16 0.00 0.00 0.00 0.00 1.00

Working with pipes

  • Using pipes is the same as any model: .fit() and .predict(), put into CVs
  • When modelling, you should spend time interrogating model predictions, plotting and printing. Does the model struggle predicting 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.