Hands-on with Feature Selection Techniques: An Introduction

Part 1: The basics of feature selection

With recent developments in big data, we’ve been given more access to high-dimensional data. Consequently, the performance of machine learning models has improved by a large margin.

On the other hand, there are significant noisy and useless features often collected or generated by different sensors and methods. These unneeded features not only influence a model’s accuracy, but they also can demand a lot of computational resources.

Therefore, feature selection is a critical process in any machine learning pipeline, designed to remove irrelevant, redundant, and noisy features and preserve a small subset of features from the primary feature space. As such, effective feature selection can help reduce computational complexity, improve model accuracy, and increase model interpretability.

With that in mind, welcome to the first post in a new series: Hands-on with Feature Selection Techniques. Throughout the series, we’ll explore a range of different methods and techniques used to select the best set of features that will help you build a simpler, faster and more reliable machine learning models.

Here’s the set of articles in this series:

This guide is intended to be a concise reference for beginners covering the most basic yet widely-used techniques for feature selection.

What is Feature Selection?

In machine learning and data science in general, Feature selection (also known as variable selection, attribute selection or subset selection) is the process by which a data scientist selects automatically or manually a subset of relevant features to use in machine learning model building.

In fact, it is one of the core concepts in machine learning which has a huge impact on your models’ performance, as it is the key to make reliable machine learning models.

Given a pool of features, the process will select the best subset of attributes that are most important and have high contribution at the time of prediction making.

Why should we select features?

It’s not always true that the more data features you have, the better the resulting model is going to be.

Data sometimes include irrelevant features that don’t result in better predictions, and sometimes redundant features that are irrelevant in the presence of others will make the learning process difficult and can result in overfitting.

Therefore, we need some techniques that will eliminate any attribute that’s could adversely affect learning.

There are several more reasons to complete feature selection, such as:

  • Simple models are easier to interpret: It’s much easier to understand the output of a model that uses 10 variables than an output that uses 100 variables.
  • Shorter training time: Reducing the number of variables reduces the computation cost, speeds up model training, and perhaps most importantly—simpler models tend to have faster prediction times.
  • Enhanced generalization by reducing overfitting: Oftentimes, many of the variables are just noise with little prediction value. However, the ML model learns from this noise and causes overfitting while simultaneously reducing generalization. By eliminating these irrelevant noisy features, we can substantially improve the generalization of ML models.
  • Variable redundancy: Features of a given dataset are frequently highly correlated, and we know that highly-correlated features provide the same information, which makes them redundant. In cases like these, we can keep just one feature and remove the redundant features without losing any information. Less redundant data means less opportunity for the model to make noise-based predictions.

Feature Selection vs. Feature Engineering

Often, newcomers to the field of machine learning may get confused between feature selection and feature engineering.

Feature engineering allows us to create new features from the ones they already have in order to help the machine learning model make more effective and accurate predictions.

Feature selection, on the other hand, allows us to select features from the feature pool (including any newly-engineered ones) that will help machine learning models more efficiently make predictions on target variables.

In a typical ML pipeline, we perform feature selection after completing feature engineering.

If you want ready-made series about feature engineering, check the links below:

Feature Selection vs. Dimensionality Reduction

Dimensionality reduction is another concept that newcomers tend to lump together with feature selection. While dimensionality reduction uses unsupervised algorithms to reduce the number of features in a dataset (as feature selection methods do), there is an important difference:

  • Feature selection is basically a process that selects and excludes some features without modifying them at all.
  • Dimensionality reduction modifies or transforms features into a lower dimension. In essence, dimensionality reduction creates a whole new feature space that looks approximately like the first one, but smaller in terms of dimensions.

The Procedure of Feature Selection

Feature selection can be described in two steps:

  • Combination of a search technique for proposing a new feature subset.
  • An evaluation measure that scores how well is the different feature subsets.

This can be computationally expensive. From all the available features, we’re looking for the best combination of feature subsets. It can be more challenging because different feature subsets render optimal performances for different machine learning algorithms.

Feature selection methods

Generally speaking, feature selection methods can be divided into three main categories:

  • Filter Methods: Rely on the features’ characteristics without using any machine learning algorithm. Very well-suited for a quick “screen and removal” of irrelevant features.
  • Wrapper methods: Consider the selection of a set of features as a search problem, then uses a predictive machine learning algorithm to select the best feature subset. In essence, these methods train a new model on each feature subset, which makes it obviously very computationally expensive. However, they provide the best performing feature subset for a given machine learning algorithm.
  • Embedded methods: Just like the wrapper methods, embedded methods take the interaction of features and models into consideration. They also perform feature selection as part of the model construction process, and they are less computationally expensive.

At the end of the series, we’ll explore more advanced techniques that use deep learning, heuristic searches, and several other ways to select features.

We’ll cover all of these methods in this series with a hands-on approach, using Python and many of its well-known data science libraries.


The reader of these articles should have some familiarity with machine learning. As mentioned above, we’ll be using Python as a programming language, as well as data science libraries like Pandas, NumPy, Matplotlib, and Scikit-learn.

You’ll need to prepare your workspace. I suggest using Anaconda since it’s pre-equipped with Python and all of the aforementioned libraries.


We’ll also be using some Kaggle datasets and other open-source datasets for demonstration purposes.

I’ve set up a GitHub repository to hold all the code sample for the series, so check it out.

You can also check the rest of the series:


Our team has been at the forefront of Artificial Intelligence and Machine Learning research for more than 15 years and we're using our collective intelligence to help others learn, understand and grow using these new technologies in ethical and sustainable ways.

Comments 0 Responses

Leave a Reply

Your email address will not be published. Required fields are marked *

wix banner square