Select with Target Mean as Performance Proxy

Method used in a KDD 2009 competition

This feature selection approach was used by data scientists at the University of Melbourne in the KDD 2009 data science competition. The task consisted in predicting churn based on a dataset with a huge number of features.

The authors describe this procedure as an aggressive non-parametric feature selection procedure that is based in contemplating the relationship between the feature and the target.

The procedure consists in the following steps:

For each categorical variable:

1) Separate into train and test

2) Determine the mean value of the target within each label of the categorical variable using the train set

3) Use that mean target value per label as the prediction (using the test set) and calculate the roc-auc.

For each numerical variable:

1) Separate into train and test

2) Divide the variable intervals

3) Calculate the mean target within each interval using the training set 

4) Use that mean target value / bin as the prediction (using the test set) and calculate the roc-auc

The authors quote the following advantages of the method:

  • Speed: computing mean and quantiles is direct and efficient
  • Stability respect to scale: extreme values for continuous variables do not skew the predictions
  • Comparable between categorical and numerical variables
  • Accommodation of non-linearities

Important The authors here use the roc-auc, but in principle, we could use any metric, including those valid for regression.

The authors sort continuous variables into percentiles, but Feature-engine gives the option to sort into equal-frequency or equal-width intervals.

Reference: Predicting customer behaviour: The University of Melbourne's KDD Cup Report. Miller et al. JMLR Workshop and Conference Proceedings 7:45-55

In [1]:
import pandas as pd
import numpy as np

from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score

from feature_engine.selection import SelectByTargetMeanPerformance
In [2]:
# load the titanic dataset
data = pd.read_csv('')

# remove unwanted variables
data.drop(labels = ['name','boat', 'ticket','body', 'home.dest'], axis=1, inplace=True)

# replace ? by Nan
data = data.replace('?', np.nan)

# missing values
data.dropna(subset=['embarked', 'fare'], inplace=True)

data['age'] = data['age'].astype('float')
data['age'] = data['age'].fillna(data['age'].mean())

data['fare'] = data['fare'].astype('float')

def get_first_cabin(row):
        return row.split()[0]
        return 'N' 
data['cabin'] = data['cabin'].apply(get_first_cabin)
In [3]:
pclass survived sex age sibsp parch fare cabin embarked
0 1 1 female 29.0000 0 0 211.3375 B5 S
1 1 1 male 0.9167 1 2 151.5500 C22 S
2 1 0 female 2.0000 1 2 151.5500 C22 S
3 1 0 male 30.0000 1 2 151.5500 C22 S
4 1 0 female 25.0000 1 2 151.5500 C22 S
In [4]:
# Variable preprocessing:

# then I will narrow down the different cabins by selecting only the
# first letter, which represents the deck in which the cabin was located

# captures first letter of string (the letter of the cabin)
data['cabin'] = data['cabin'].str[0]

# now we will rename those cabin letters that appear only 1 or 2 in the
# dataset by N

# replace rare cabins by N
data['cabin'] = np.where(data['cabin'].isin(['T', 'G']), 'N', data['cabin'])

array(['B', 'C', 'E', 'D', 'A', 'N', 'F'], dtype=object)
In [5]:
pclass        int64
survived      int64
sex          object
age         float64
sibsp         int64
parch         int64
fare        float64
cabin        object
embarked     object
dtype: object
In [6]:
# number of passengers per value
0    999
1    170
2    113
3      8
5      6
4      6
9      2
6      2
Name: parch, dtype: int64
In [7]:
# cap variable at 3, the rest of the values are
# shown by too few observations

data['parch'] = np.where(data['parch']>3,3,data['parch'])
In [8]:
0    888
1    319
2     42
4     22
3     20
8      9
5      6
Name: sibsp, dtype: int64
In [9]:
# cap variable at 3, the rest of the values are
# shown by too few observations

data['sibsp'] = np.where(data['sibsp']>3,3,data['sibsp'])
In [10]:
# cast discrete variables as categorical

# feature-engine considers categorical variables all those of type
# object. So in order to work with numerical variables as if they
# were categorical, we  need to cast them as object

data[['pclass','sibsp','parch']] = data[['pclass','sibsp','parch']].astype('O')
In [11]:
# check absence of missing data

pclass      0
survived    0
sex         0
age         0
sibsp       0
parch       0
fare        0
cabin       0
embarked    0
dtype: int64


In all feature selection procedures, it is good practice to select the features by examining only the training set. And this is to avoid overfit.

In [12]:
# separate train and test sets

X_train, X_test, y_train, y_test = train_test_split(
    data.drop(['survived'], axis=1),

X_train.shape, X_test.shape
((914, 8), (392, 8))
In [13]:
# feautre engine automates the selection for both
# categorical and numerical variables

sel = SelectByTargetMeanPerformance(
    variables=None, # automatically finds categorical and numerical variables
    scoring="roc_auc_score", # the metric to evaluate performance
    threshold=0.6, # the threshold for feature selection, 
    bins=3, # the number of intervals to discretise the numerical variables
    strategy="equal_frequency", # whether the intervals should be of equal size or equal number of observations
    cv=2,# cross validation
    random_state=1, #seed for reproducibility
), y_train)
SelectByTargetMeanPerformance(bins=3, cv=2, random_state=1,
                              strategy='equal_frequency', threshold=0.6,
                              variables=['pclass', 'sex', 'age', 'sibsp',
                                         'parch', 'fare', 'cabin', 'embarked'])
In [14]:
# after fitting, we can find the categorical variables
# using this attribute

['pclass', 'sex', 'sibsp', 'parch', 'cabin', 'embarked']
In [15]:
# and here we find the numerical variables

['age', 'fare']
In [16]:
# here the selector stores the roc-auc per feature

{'pclass': 0.6802934787230475,
 'sex': 0.7491365252482871,
 'age': 0.5345141148737766,
 'sibsp': 0.5720480307315783,
 'parch': 0.5243557188989476,
 'fare': 0.6600883312700917,
 'cabin': 0.6379782658154696,
 'embarked': 0.5672382248783936}
In [17]:
# and these are the features that will be dropped

['age', 'sibsp', 'parch', 'embarked']
In [18]:
X_train = sel.transform(X_train)
X_test = sel.transform(X_test)

X_train.shape, X_test.shape
((914, 4), (392, 4))

That is all for this lecture, I hope you enjoyed it and see you in the next one!

In [ ]: