Dimensionality reduction is the process of transforming high-dimensional data into a lower-dimensional format while preserving its most important properties. This technique has applications in many industries including quantitative finance, healthcare, and drug discovery. The applications of dimensionality reduction are numerous, so every data scientist should know some state-of-the-art methods for carrying it out. 

An easy-to-understand method for dimensionality reduction is random forests feature importance. Random forests is a machine learning algorithm that uses many decision trees for classification and regression tasks. Each decision tree asks yes or no questions about the values in the data and successively splits them into similar groups based on the answer. Most random forest packages allow you to then pull important features, which are the fields that most effectively separate distinct parts of the data. 

For example, when predicting whether or not a borrower will default on a loan, credit score would be an important feature that would most effectively split the data into defaulting and non-defaulting borrowers. Although this method is useful, it is limited to supervised learning processes in which we know the output we’re trying to predict. In this example, that output is whether a borrower will default or not. 

In some use cases, data scientists have data but don’t have a framed supervised learning problem. Supervised learning requires labeled data which means for every input we have a label that represents the outcome we are trying to predict. Similarly, use cases also exist in which a data scientist has access to high-dimensional, unlabeled data which simply lacks a field that specifies an output for a predictive model. An example is having access to loan borrower data without any information on loan status (i.e., default/no default). 

When faced with the issue of high-dimensional, unlabeled data (e.g., hundreds to thousands of columns), you can employ unsupervised dimensionality reduction techniques. One of the most common unsupervised learning methods of dimensionality reduction is principal component analysis (PCA). PCA represents a data set with a large number of columns with a smaller data set with fewer columns, which are called principal components. These can then be used to analyze trends, clusters, and outliers and can even help frame a supervised learning problem. 

A video on dimensionality reduction techniques.

Scikit-learn is a Python machine learning library that has many easy-to-use modules to carry out dimensionality reduction. The ensemble module in Scikit-learn has random forest algorithms for both classification and regression tasks. In each of the supervised learning use cases, random forest can be used to reduce the number of dimensions in data. For unsupervised dimensionality reduction tasks, the decomposition module has packages for PCA. 

Here, we will apply random forest feature selection and PCA to the Lending Club data set, which is from a peer-to-peer lending company that released its data to the public. The data set contains borrower credit history and loan statuses.

What Is Dimensionality Reduction?

Dimensionality reduction is the process of transforming high-dimensional data into a lower-dimensional format while preserving its most important properties. This technique has applications in many industries including quantitative finance, healthcare, and drug discovery.

More From Sadrach PierreA Guide to Data Clustering Methods in Python

Reading in and Preparing the Data 

Let’s start by importing the Pandas library:

import pandas as pd 

We then can relax the display limits for rows and columns:

pd.set_option('display.max_columns', None)
pd.set_option('display.max_rows', None)

Next, let’s read our data into a Pandas data frame:

df = pd.read_csv()

Let’s see how many columns and rows are in our data:

print("Number of Columns: ", len(list(df.columns)))
print("Number of rows: ", len(df))
A data readout
Image: Screenshot by the author.

We see that we have 142 columns and 2.9 million rows. This is a pretty large data set to work with on most machines, so it’s a great candidate for dimensionality reduction. 

Now, let’s display the first five rows of the data. Since this data set is relatively large, the columns are truncated in the screenshot below. When you run this code locally on your machine you will be able to see all of the columns:

A screenshot of a data readout
Image: Screenshot by the author

Working with a data set of this size on a local machine can be cumbersome. As you’ll notice when you run this code, simple tasks such as reading in the data and displaying take quite some time. Thus, we’ll just work with a subset of the data here. We’ll only consider credit card repayment loans in our analysis. This corresponds to the purpose column with the value of credit_card. Let’s filter our data to only have credit card repayment loans:

df = df[df['purpose'] == 'credit_card']

Let’s also take a small subset of the columns. We’ll consider the following fields in the data:

columns = ['loan_amnt', 'loan_status','funded_amnt', 
'funded_amnt_inv', 'term', 
 'num_accts_ever_120_pd', 'num_actv_bc_tl', 
'num_actv_rev_tl', 'avg_cur_bal', 'bc_open_to_buy', 'bc_util', 
'chargeoff_within_12_mths', 'delinq_amnt', 'last_fico_range_low', 

df = df[columns]

For details on what each of these fields mean, see the data dictionary. Some important fields are loan amount, interest rates, home ownership status, FICO scores, and number of active accounts. 

Now, let’s also write the filtered data frame to a new csv file that we will call credit_card_loans.csv:

df.to_csv("credit_card_loan.csv", index=False)

Now, let’s read in our new csv file into a separate data frame. We will call this new dataframe df_credit:

df_credit = pd.read_csv("credit_card_loan.csv")

Now, let’s print the new number of rows and columns:

print("Number of Columns: ", len(list(df_credit.columns)))
print("Number of rows: ", len(df_credit))
A screenshot of a data readout
Image: Screenshot by the author.

We see that our data now has 18 columns and 695,665 rows. This size set is significantly easier to work with. 

Next, let’s print the first five rows of data:

A screenshot of a data readout
Image: Screenshot by the author.

Before we discuss any specific methods, note that we were already able to significantly reduce the dimensionality of the data in terms of both columns and rows. But we still need to do a bit more data prep. Notice some columns have missing values — NaN means not a number. Let’s impute those with the mean for each of these columns:

def fill_na(numerical_column):       
df_credit[numerical_column].fillna(df_credit[numerical_column].mean(), inplace=True)

A screenshot of a data readout
Image: Screenshot by the author.

We see the missing values have been imputed. Next, let’s convert the categorical columns into codes that are machine-readable. This conversion is necessary when using most machine learning packages available in Python:

A screenshot of a data readout
Image: Screenshot by the author.

Now, the home ownership, term, and verification status columns have corresponding categorical columns. The first method we’ll look at is using random forests feature importance to reduce dimensionality. This is a supervised machine learning method since random forests require labeled data. 

So, the next thing we need to do is generate labels from the loan_status columns. First, let’s print the unique set of values for loan status:

A screenshot of a data readout
Image: Screenshot by the author.

For simplicity, let’s only consider the loan status outcomes fully paid and default/charged off. We will also combine these. 

df_credit = df_credit[df_credit['loan_status'].isin(['Fully Paid', 
'Default', 'Charged Off'])]

Let’s also create binary labels for these loan status outcomes. A value of one will correspond to default/charged off, meaning the loan wasn’t paid off and has gone into collections, and zero means the loan was fully paid off:

df_credit['loan_status_label'] = np.where(df_credit['loan_status'] == 
'Fully Paid', 0, 1)
A screenshot of a data readout
Image: Screenshot by the author.

Finally, let’s filter the columns in our data frame so we only have columns with machine-readable values:

columns2 = ['loan_amnt', 'loan_status_label', 'funded_amnt', 
'funded_amnt_inv', 'term_cat', 
 'num_accts_ever_120_pd', 'num_actv_bc_tl', 'num_actv_rev_tl', 
'avg_cur_bal', 'bc_open_to_buy', 'bc_util', 
'chargeoff_within_12_mths', 'delinq_amnt', 'last_fico_range_low', 
df_credit = df_credit[columns2]
A screenshot of a data readout
Image: Screenshot by the author.

Finally, let’s convert the interest rate column into a numerical column:

df_credit['int_rate'] = df_credit['int_rate'].str.rstrip('%')
df_credit['int_rate'] = df_credit['int_rate'].astype(float)
df_credit.fillna(0, inplace=True)


Random Forests

We are now in good shape to apply random forests, which is a tree-based ensemble algorithm that constructs a series of tree data structures and asks yes-or-no questions about the statistics in the data. Each of these trees makes a prediction based on the answers, and the trees are combined to make a single prediction. Let’s import the random forest classifier and the train/test split methods from Scikit-learn

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

Let’s define our input and output and split our data for training and testing. This step is necessary so that we don’t overfit to noise in our training data and ensures that our model can generalize well when we make predictions on future data. 

X = df_credit[['loan_amnt', 'funded_amnt', 'funded_amnt_inv', 
'term_cat', 'int_rate','mths_since_recent_revol_delinq','home_ownership_cat', 'verification_status_cat',
 'num_accts_ever_120_pd', 'num_actv_bc_tl', 'num_actv_rev_tl', 
'avg_cur_bal', 'bc_open_to_buy', 'bc_util', 
'chargeoff_within_12_mths', 'delinq_amnt', 'last_fico_range_low', 
y = df_credit['loan_status_label']

X_train, X_test, y_train, y_test = train_test_split(X, y , 
random_state=42, test_size = 0.33)

Next, let’s fit our random forest model to the training data and generate a plot for feature importance:

import seaborn as sns
import matplotlib.pyplot as plt
model = RandomForestClassifier()
model.fit(X_train, y_train)

features = ['loan_amnt', 'funded_amnt', 'funded_amnt_inv', 'term_cat', 'int_rate','mths_since_recent_revol_delinq','home_ownership_cat', 
 'num_accts_ever_120_pd', 'num_actv_bc_tl', 'num_actv_rev_tl', 
'avg_cur_bal', 'bc_open_to_buy', 'bc_util', 
'chargeoff_within_12_mths', 'delinq_amnt', 'last_fico_range_low', 
feature_df = pd.DataFrame({"Importance":model.feature_importances_, 
"Features": features })
plt.bar(feature_df["Features"], feature_df["Importance"])
plt.title("Random Forest Model Feature Importance")
A graphic plot of the data
Image: Screenshot by the author. 

We see that the FICO scores, interest rate, and current balance are the three most important features. We can clearly use the random forest feature importance to narrow down which factors to consider when predicting credit risk. 

The downside with this method is that it assumes we have labels. Here, it is loan status. Often, though, we can find use cases in which we’d like to narrow down a large list of columns in data without having any labels. The most common technique for this approach is principal component analysis. 


Principal Component Analysis 

Principal component analysis works by finding a smaller set of column values from an uncorrelated larger set. This method works by representing independent, uncorrelated features as a sum of the original features. 

Let’s start by importing the PCA package from Sklearn. We will also need the StandardScaler method from the preprocessing module in Sklearn.

from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler

Now, let’s define the input for our PCA algorithm:

X = df_credit[features2]

Next, let’s scale our data using the standardScaler method. This step helps with numerical stability when the algorithm computes the components:

scaler = StandardScaler()
X_scaled = scaler.fit(X)

Next, let’s define a PCA object with four components, fit to X_scaled, then generate our components:


We can then store the component in a Pandas data frame:

components_df = pd.DataFrame({'component_one': 
list(X_components[:,0]), 'component_two': list(X_components[:,1]),
list(X_components[:,2]), 'component_four': list(X_components[:,3])})

A screenshot of a data readout
Image: Screenshot by the author.

Now, we’ll store the class in a variable called labels and define some variables that we’ll use for formatting our plot:


We can now generate our scatter plot and add axes labels and a title. We will look at a scatter plot of the first two components:

for i in np.unique(labels):   
index=np.where(labels==i)   ax.scatter(components_df['component_one'].loc[index],components_df['component_two'].loc[index],c=color_dict[i],s=10,
plt.xlabel("1st Component",fontsize=14)
plt.ylabel("2nd Component",fontsize=14)
plt.title('Scatter plot of Principal Components')
A scatter plot of the data
Image: Screenshot by the author.

We can see a distinct separation between classes in our scatter plot. Although we have labels that we can use for a sanity check for our PCA algorithm, in practice, PCA is used in an unsupervised manner. This means we can use these methods to explore distinct clusters for any column we are interested in. 

For example, we can explore clusters in credit scores or even income. Cluster analysis of these values may be useful for lenders because it allows them to consider creditworthy borrowers who would otherwise be denied a loan. An example is a recent graduate student with a low credit score who recently got a job in the tech industry with a high-six-figure salary. While the borrower was in graduate school, they may have had trouble paying off loans, but they may now be creditworthy upon starting their new job. 

This type of cluster analysis can help detect these borrowers who would otherwise be denied for a loan which can translate to higher revenue for lenders. PCA is very powerful in this sense since it can allow you to analyze distinct groups in your data without requiring any predefined labels.

The code in this post is available on GitHub

More in Data ScienceA Primer on Model Fitting

The Takeaway

Most data science teams across industries are faced with the task of reducing the dimensionality of data. This may be for the sake of simple analytics, building interpretable models, or even performing cluster analysis on a large data set. Random forest is useful for dimensionality reduction when you have a well-defined supervised learning problem. In our case, our labels were credit default and fully paid off loans. Random forest is also appealing because you can directly interpret how significant the features used in your model are for determining an outcome. PCA is a powerful tool when you do not have labels in your data. Companies can use PCA to explore distinct groups in data which can then be used to make decisions. 

Expert Contributors

Built In’s expert contributor network publishes thoughtful, solutions-oriented stories written by innovative tech professionals. It is the tech industry’s definitive destination for sharing compelling, first-person accounts of problem-solving on the road to innovation.

Learn More

Great Companies Need Great People. That's Where We Come In.

Recruit With Us