Wrapper custom class for scikit-learn's Iterative Imputer for use with cross_val_score()

Question

Scikit-learn's Iterative Imputer can impute missing values in a round-robin fashion. To evaluate its performance against other conventional regressors, it is possible to build a simple pipeline and get scoring metrics from cross_val_score. The issue is that Iterative Imputer does not have a 'predict' method as per error:

AttributeError: 'IterativeImputer' object has no attribute 'predict'

See a minimum example of what is trying to be achieved:

# import libraries
import pandas as pd
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import Pipeline

# define scaler, model and pipeline
scaler = StandardScaler() # use any scaler
imputer = IterativeImputer() # with any estimator, default = BayesianRidge()
pipeline = Pipeline(steps=[('s', scaler), ('i', imputer)])

train, test = df.values, df['A'].values 
scores = cross_val_score(pipeline, train, test, cv=10, scoring='r2')
print(scores)

What possible solutions exist? If a custom wrapper is needed, how should it be written to include the 'predict' method?

Answer 1

cross_val_score needs pipeline with model at the end (which has predict)

scaler  = StandardScaler()
imputer = IterativeImputer()
model   = BayesianRidge()  # any model

pipeline = Pipeline(steps=[('s', scaler), ('i', imputer), ('m', model)])

cross_val_score without model make no sense.

---

I see also other problem - with values train, test which you use in cross_val_score.

It should be X, y instead of train, test but it is only names so it is not so importalt but important is what you assing to variables.

Problem is that X should be without y but you use train = df.values so you create X with y

df_train = pd.DataFrame({
                'X': range(20), 
                'y': range(20),
           })

X_train = df_train[ ['X'] ]  # it needs inner `[]` to create DataFrame, not Series
y_train = df_train[  'y'  ]  # it has to be single column (Series)

scores = cross_val_score(pipeline, X_train, y_train, cv=10, scoring='r2')

(BTW: you don't have to use .values)

The same with more columns

df_train = pd.DataFrame({
                'A': range(20), 
                'B': range(20), 
                'y': range(20),
           })

X_train = df_train[ ['A', 'B'] ]
y_train = df_train[ 'y' ]

---

Minimal working code but with fake data (which are useless)

# import libraries
import pandas as pd
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import Pipeline
from sklearn.linear_model import BayesianRidge

df_train = pd.DataFrame({
                'A': range(100),  # fake data
                'B': range(100),  # fake data
                'y': range(100),  # fake data
           })

df_test = pd.DataFrame({
                'A': range(20),  # fake data
                'B': range(20),  # fake data
                'y': range(20),  # fake data
           })

# define scaler, model and pipeline
scaler  = StandardScaler()
imputer = IterativeImputer()
model   = BayesianRidge()

pipeline = Pipeline(steps=[('s', scaler), ('i', imputer), ('m', model)])

X_train = df_train[ ['A', 'B'] ]  # it needs inner `[]` to create DataFrame, not Series
y_train = df_train[ 'y' ]         # it has to be single column (Series)

scores = cross_val_score(pipeline, X_train, y_train, cv=10, scoring='r2')
print(scores)

X_test = df_test[['A', 'B']]
y_test = df_test['y']

scores = cross_val_score(pipeline, X_test, y_test, cv=10, scoring='r2')
print(scores)