Feature Engineering: The Ultimate Guide

As a data enthusiast, feature engineering is one of the key concepts that is essential to easily manage your data and present it perfectly as needed. In most cases, feature engineering is associated to data science as it is used to improve performance of the models that are being trained. However, feature engineering is not only restricted to data science, it has also been used in data analysis task. This article is an ultimate guide to those in the data field and those seeking to get into the data field.

what is Feature Engineering?

Feature engineering refers to a set of processes involving data manipulation of rows and columns, conducted during data preprocessing to come up with new meaningful features.
Data science and machine learning is all about data fed into the models. The outcome we get directly depends on the quality of data and features we give it. This is why there is a need to transform the raw data into an understandable format by giving an easy and understandable format that shows the underlying relationships and patterns within the data.

Tools for feature Engineering
Python
It offers a number of libraries that can be used in the process. The most commonly used libraries are:
pandas
Sklearn

Importance of feature engineering
Feature engineering presents a huge number of advantages. Some of the most important features that it presents are shown in the below list.

Improves accuracy of a model. The feature engineering process involves removing a huge set of irrelevant information within a given dataset. For example, you can get a feature that combines both date and time on the same column. Feature engineering involves capturing the underlying pattern of the column, separating the time and the date. This increases the accuracy of the model because it allows the model get to know the differences training more effectively.
Highlights the key insights. creating new features gives more information about the data which can help in training a model or understanding a model more effectively. For example, a dataset may contain a feature such as total cost and cost per unit. Through feature engineering, a new feature can be brought up to show the number of units. This can be of help during visualization of the data and even training of models. Overally, when feature engineering activities are done correctly, the resulting data set is optimal and contains all of the important factors that affect the business problem.

Techniques involved in feature engineering
There are a number of techniques that can be used, depending on the types of features in the dataset. In most cases, the data are either in a categorical, numerical or in a text format. The below are some of the main features engineering techniques that one must be familiar with when preparing data for any data science or machine learning process.

1. Imputation
This technique is mostly used when handling missing values. It basically is filling missing values in a dataset. Missing values is one of the most common problems faced when dealing with large datasets. To get the desired results, the missing values must be handled to make the model learn perfectly.
For Numerical features, Numerical imputation is how to handle the missing values. There are different ways of achieving this, and it depends on the distribution of the feature. It can be done with the process such as calculating the median, mean, mode, or just simply filling the fields with a null value. Python has functions such as the fillna() which is mostly used for this process.
Categorical features are handled differently because performing mathematical operations is not possible. Counting the number of appearance of each category is the first process necessary in determining the number of appearance of each and every category. The missing categories can then be filled based on the number of appearance of the categories. For example, the most appearing category can be used to fill the missing categories. In other cases where there is a large number of missing values, you might create a new category for example, "other" and assign it.

2. Feature scaling
For models to perfectly predict, it is important for the features to be in the same range and it a standard format. Its mostly considered by individuals as one of the most difficult activities involved in predicting features. Most datasets have different values, some features may be continuous values, others may be fixed values. After the process of scaling, the features then become the same depending on the scaling technique that is used. The most essential and common ways of scaling widely used are:
a. Normalization. This scaling technique involves converting all the values to be in a range of 0 and 1. Before using this scaling technique, it is mostly adviced to deal with the outliers in the dataset to avoid showing some slight difference in the scaled features.
b. Standardization. This scaling technique involves setting a mean of 0 and a standard deviation of 1. From the sklearn library in python, you can import the StandardScaler function which is used to scale large number of datasets.

3. Encoding categories
The data science models cannot transform the categorical or textual data that is passed to it. During the feature engineering process, the categorical data should be first converted to a format that the machines can learn. This is the use of the encoding technique, depending on the feature. The below are the most used encoding techniques.
a. Label Encoding. This involves giving each and every column a unique integer to represent it. However, some models might treat the values assigned to represent the categories in an order. For example when encoding the gender and male is set to 1 then female to 0, the models might take 1 to be higher than 0 instead of treating them both the same.
b. One hot encoding. A feature that creates each column for categorical values. For example when dealing with the gender, when encoding it it adds a different column and if the row is male, it sets the male to 1 and the female column to 0. This provides the model with additional information that can be used in the learning process of data science and ML models.

4. Feature Transformation
Datasets also come in different states. Some may be skewed, non-linear, and outliers. Feature transformation involves converting the data to have linear relationships between columns, converting to a less skewed or non skewed, making the distribution to be normal and even removing the outliers. The technique involves processes such as Log Transform, and Reciprocal transformation.

Feature Engineering Ultimate Guide Summary
Feature engineering is an important role that each and every data scientist must understand. A step by step guide guide can help get the best out of the training models when a user inserts the feature. The below is a summary of how you can go about it.