Hackernoon logoImprove Machine Learning Model Performance by Combining Categorical Features by@davisdavid

Improve Machine Learning Model Performance by Combining Categorical Features

Categorical features are types of data that may be divided into groups. They are three common categorical data types: binary, binary and nominal. In this article, we will learn how combining categorical features can improve your machine learning model performance. We are going to use the Financial Inclusion in Africa dataset from the Zindi competition page. The objective of this dataset is to predict who is most likely to have a bank account. We have 13 variables in the dataset, 12 independent variables, and 1 dependent variable. The next step is to separate the variables and target(bank_account) from the data.
image
Davis David Hacker Noon profile picture

@davisdavidDavis David

Data Scientist | AI Practitioner | Software Developer. Giving talks, teaching, writing.

When you training a machine learning model, you can have some features in your dataset that represent categorical values. Categorical features are types of data that may be divided into groups. 

There are three common categorical data types which are:

  1. Ordinal - This has a set of orders. Example: rating happiness on a scale of 1-10
  2. Binary - This has only two values. Example: Male or Female
  3. Nominal - This does not have any set of orders. Example: Countries  

Most machine learning algorithms require numerical input and output variables. Therefore you will have to transform categorical features in your dataset into integers or floats to be utilized by machine learning algorithms. You can either use LabelEncoding for the binary features or the One-hot-encoding method for nominal features. 

In this article, you will learn how combining categorical features can improve your machine learning model performance.

So let’s get started.🚀

Combining Categorical Features in Machine Learning Models

You can create a new feature that is a combination of the other two categorical features. You can also combine more than three or four or even more categorical features.

df["new_feature"] = (
	df.feature_1.astype(str)
	 + "_"
	 + df.feature_2.astype(str)
	)

In the above code, you can see how you can combine two categorical features by using pandas and form a new feature in your dataset.

So which categorical features should you combine? Well, there isn't an easy answer to that. It depends on your data and the types of features. Some domain knowledge might be useful for creating new features like this.

To illustrate the whole process, we are going to use the Financial Inclusion in Africa dataset from the  Zindi competition page, which has many categorical features that we can combine some of these features and see if we can improve the model performance.

The objective of this dataset is to predict who is most likely to have a bank account. So this is a classification problem.

1.Load the dataset

Our first step is to make sure that we have downloaded the dataset provided in the competition. You can download the dataset here.

Import important python packages.

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import warnings
np.random.seed(123)
warnings.filterwarnings('ignore')
%matplotlib inline

Load the dataset.

# Import data
data = pd.read_csv('data/Train_v2.csv')

Let’s observe the shape of our dataset.

# print shape
print('data shape :', data.shape)

data shape : (23524, 13)

The above output shows the number of rows and columns in the dataset. We have 13 variables in the dataset, 12 independent variables, and 1 dependent variable. 

We can observe the first five rows from our data set by using the head() method from the pandas library.

# inspect data 

data.head()
image

It is important to understand the meaning of each feature so you can really understand the dataset. You can read the VariableDefinition.csv file to understand the meaning of each variable presented in the dataset.

2.Understand The Dataset

We can get more information about the features presented by using the info() method from pandas.

#show Some information about the dataset

print(train_data.info())
image

The output shows the list of variables/features, sizes if it contains missing values, and data type for each variable. From the dataset, we don’t have any missing values and we have 3 features of integer data type and 10 features of the object data type(most are categorical features).

3. Data preparation for machine learning models

The next step is to separate the independent variables and target(bank_account) from the data. Then transform the target values from the object data type into numerical by using LabelEncoder.

#import preprocessing module
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import MinMaxScaler

# Convert target label to numerical Data
le = LabelEncoder()
data['bank_account'] = le.fit_transform(data['bank_account'])

#Separate training features from target
X = data.drop(['bank_account'], axis=1)
y = data['bank_account']

print(y)
image

The target values have been transformed into numerical data types, 1 represents ‘Yes’ and 0 represents ‘No’.

I have created a simple preprocessing function to:

# function to preprocess our data 

def preprocessing_data(data):

    # Convert the following numerical labels from interger to float
    float_array = data[["household_size", "age_of_respondent", "year"]].values.astype(float
    )
    
    # categorical features to be converted to One Hot Encoding
    categ = [
        "relationship_with_head",
        "marital_status",
        "education_level",
        "job_type",
        "country",
    ]
    
    # One Hot Encoding conversion
    data = pd.get_dummies(data, prefix_sep="_", columns=categ)
    
    # Label Encoder conversion
    data["location_type"] = le.fit_transform(data["location_type"])
    data["cellphone_access"] = le.fit_transform(data["cellphone_access"])
    data["gender_of_respondent"] = le.fit_transform(data["gender_of_respondent"])
    
    # drop uniquid column
    data = data.drop(["uniquid"]), axis=1)
    
    # scale our data 
    scaler = StandardScaler()
    data = scaler.fit_transform(data)
    
    return data

Let’s preprocess our dataset.

# preprocess the train data 
processed_test_data = preprocessing_data(X_train)

4. Model Building and Experiments

A portion of the data set will be used to evaluate our models.

# Split train_data
from sklearn.model_selection import train_test_spilt
X_Train, X_val, y_Train, y_val = train_test_split(processed_train_data, y_train, stratify = y, test_size = 0.1, random_state=42)

Only 10% of the dataset will be used for evaluating the machine learning models. The parameter stratify = y will ensure an equal balance of values from both classes (‘yes’ and ‘no’) for both train and validation set.

We will use the Logistic Regression algorithm for this classification problem to train and predict who is most likely to have a bank account.

#import classifier algorithm here
from sklearn.linear_model import LogisticRegression

# create classifier
lg_model = LogisticRegression()

#Training the classifier
lg_model.fit(X_Train,y_Train)

After training the classifier, let’s use the trained model to predict our evaluation set and see how it performs. We will use Accuracy as our evaluation metric.

# import evaluation metrics
from sklearn.metrics import confusion_matrix, accuracy_score

# evaluate the model
y_pred = lg_model.predict(X_val)

# Get the accuracy
print("Accuracy Score of Logistic Regression classifier: ","{:.4f}".format(accuracy_score(y_val, lg_y_pred)))

Accuracy Score of Logistic Regression classifier:  0.8874 

1st Experiment: Combine education_level and job_type features.

Now that we know the basic model performance, let’s see if we can improve it by combining the education_level and job_type features.

What we need to do in our first experiment is to update the preprocessing function we have created and then run the rest of the code.

# function to preprocess our data 
 
def preprocessing_data(data):

    # Convert the following numerical labels from integer to float
    float_array = data[["household_size", "age_of_respondent", "year"]].values.astype(float)

    # combine some cat features 
    data["features_combination"] = (data.education_level.astype(str) + "_" + data.job_type.astype(str) )

    # remove individual features that are combined together
    data = data.drop(['education_level','job_type'], axis=1)

    # categorical features to be converted by One Hot Encoding
    categ = [
      "relationship_with_head",
      "marital_status",
      "features_combination",
      "country"
      ]

    # One Hot Encoding conversion
    data = pd.get_dummies(data, prefix_sep="_", columns=categ)

    # Label Encoder conversion
    data["location_type"] = le.fit_transform(data["location_type"])
    data["cellphone_access"] = le.fit_transform(data["cellphone_access"])
    data["gender_of_respondent"] = le.fit_transform(data["gender_of_respondent"])

    # drop uniquid column
    data = data.drop(["uniqueid"], axis=1)

    # scale our data 
    scaler = StandardScaler()
    data = scaler.fit_transform(data)

    return data

In the above preprocessing function I have updated the code by

  • Combine educaion_level and job_type to create a new feature called “features_combination”.
  • Remove individual features (education_level and job_type) from the dataset.
  • Add a new feature called “feature_combinaton” in the list of categorical features to be converted by One Hot Encoding.

Note: I have selected only Nominal categorical features (have more than 2 unique values).

After retraining the logistic regression classifier for the 1st experiment, the model performance increased from 0.8874 to 0.8882. This show that combining categorical feature can improve the model performance.
Keep in mind that we did not change anything such as hyper-parameters in your machine learning classifier.

2nd Experiment: Combine relation_with_head and marital_status features

In our second experiment, we are going to combine the other two categorical features which are relationship_with_head and marital_status.

We just need to update the preprocessing function (like in the 1st experiment) and then run the rest of the code.

# function to preprocess our data 

def preprocessing_data(data):

    # Convert the following numerical labels from integer to float
    float_array = data[["household_size", "age_of_respondent", "year"]].values.astype(
        float
    )
    
    # combine some cat features 
    data["features_combination"] = (data.relationship_with_head.astype(str) + "_"
                           + data.marital_status.astype(str) 
                      )
    # remove individual features that are combined together
    data = data.drop(['relationship_with_head','marital_status'], axis=1)


    # categorical features to be converted by One Hot Encoding
    categ = [
        "features_combination",
        "education_level",
        "job_type",
        "country",
    ]

    # One Hot Encoding conversion
    data = pd.get_dummies(data, prefix_sep="_", columns=categ)

    # Label Encoder conversion
    data["location_type"] = le.fit_transform(data["location_type"])
    data["cellphone_access"] = le.fit_transform(data["cellphone_access"])
    data["gender_of_respondent"] = le.fit_transform(data["gender_of_respondent"])

    # drop uniquid column
    data = data.drop(["uniqueid"], axis=1)

    # scale our data 
    scaler = StandardScaler()
    data = scaler.fit_transform(data)

    return data

In the above preprocessing function I have updated the code by

  • Combine relation_with_head and marital_status to create a new feature called “features_combination”.
  • Remove individual features (relation_with_head and marital_status) from the dataset.
  •  Add a new feature called “feature_combination” in the list of categorical features to be converted by One Hot Encoding.

After retraining the logistic regression classifier for the 2nd experiment, the model performance decreased from 0.8874 to 0.8865.

This shows that sometimes when you combine categorical features your machine learning model will not improve as you expected. Therefore you will need to run a lot of experiments until you get satisfactory performance from your machine learning model.

Wrapping Up

In this article, you have learned how to combine categorical features in your dataset in order to improve the performance of your machine learning model. As I have said to get satisfactory performance for your model, you need to have domain knowledge about the problem you are solving. Also, you need to run a lot of experiments that require more computational resources.

Congratulations 👏👏, you have made it to the end of this article! I hope you have learned something new that will help you on your next machine learning or data science project.

If you learned something new or enjoyed reading this article, please share it so that others can see it. Until then, see you in the next post!

You can also find me on Twitter @Davis_McDavid.

And you can read more articles like this here.

Tags

Join Hacker Noon

Create your free account to unlock your custom reading experience.