Feature engineering

Feature engineering is a procedure in machine learning and data mining to prepare data into a form suitable for machine learning algorithms. It is also described as feature vectorization, data transformation, or feature construction. The purpose of feature engineering is to select significant features from raw data and then manipulate it into a form suitable for machine learning or predictive modelling.

The feature engineering process involves transforming data into numerical representation, such as quantifying qualitative attributes or standardising numerical features. A variety of techniques are commonly used, including discretisation, normalisation, aggregation, selection, grouping, and transforming.

Discretisation involves splitting continuous features into discrete values, typically known as bins. This reduces the probability of overfitting and simplifies feature relationships.

Normalisation is the process of scaling features to a range that is appropriate for the particular algorithm used.

Aggregation carries out mathematical operations such as summation, averages, minimum or maximum values over relevant instances of a feature.

Selection typically involves reducing the list of features of interest to a smaller subset.

Grouping is a similar technique but involves combining similar features and then applying aggregation.

Transforming involves creating new features from existing features, which may be non characteristic for the machine learning ability, but provide meaningful insight into the data.

Feature engineering is an important part of the process in many machine learning tasks as it can lead to improved accuracy. The process of feature engineering requires an understanding of machine learning algorithms and the data, and the ability to select and transform features in a way that enhances the performance. As the success of a machine learning algorithm often relies heavily on the feature engineering process, it has become an indispensable part in many machine learning tasks.