What does Feature Engineering mean?

Feature engineering refers to turning raw data into features a model can actually work with. That includes reshaping existing values, combining them, bringing them onto a common scale, or removing unnecessary columns. The step sits between data provisioning and training, and in classic methods it decides the quality of the result more than the choice of algorithm does.

Typical operations include normalizing to a common value range, one-hot encoding categorical fields, and splitting a date into weekday, month, and holiday flag. Derived quantities emerge from existing columns, such as the ratio between two amounts or the distance to last year’s value. Feature selection then removes columns that contribute nothing, to reduce noise and computing time. With deep learning, this step largely disappears, because the network forms its own features.

Feature engineering pays off wherever domain knowledge about the data exists that a model cannot derive from raw columns on its own. Examples include financial ratios in accounting, quantities derived from sensor readings, and temporal patterns in booking and order data. For images, audio, and free text, deep networks take over this work instead.

The advantage over a model built on every available column is stability. A few well-justified features deliver more reliable results than many randomly correlating ones. Because every feature is named and described, the preparation can also be applied identically to new data later on.

Two mistakes occur regularly: features that are not yet available at the time of prediction, and features that indirectly encode protected characteristics such as origin or gender. Both need to be checked before training, since they are barely detectable once the model is running.