Feature Engineering
Transforming raw data into informative inputs that make a learning problem easier to solve.
Definition
Feature engineering is the craft of turning raw measurements into representations a model can use effectively: scaling, combining, encoding, and deriving new quantities from domain knowledge.
Feature engineering and model choice interact: tree ensembles handle raw, unscaled features and interactions well, while linear models and neural networks benefit from careful scaling and encoding. Knowing which burden falls on the features and which on the model is part of the craft.
Automated feature learning has reduced but not eliminated the need for human insight. Deep networks learn features from raw data, yet the way inputs are represented and normalized still shapes what they can learn, and in tabular and scientific settings, physically meaningful derived quantities often outperform anything a generic model discovers on its own from limited data.
Typical operations
- Normalization and standardization of numeric ranges.
- One-hot encoding of categorical variables.
- Deriving ratios, differences, or physically meaningful quantities.
- Log or power transforms to tame skewed distributions.
Why it matters
Good features often matter more than the choice of algorithm. Encoding domain knowledge into inputs can turn an intractable problem into an easy one. Deep learning reduces but does not eliminate this work, since raw representation still affects what a network can learn.
Fusion connection
Kronos surrogates often take physically dimensionless groups as inputs rather than raw machine settings, because such features generalize better across the Hyperion design space.