Feature Engineering

The biggest gains in a model often come not from a smarter algorithm, but from better inputs. Feature engineering is the craft of turning raw columns into variables that actually carry signal.

High-Value Transformations

Ratios and differences

Combine two columns into one that captures a rate or a gap: revenue per user, price per gram, days since last order.

Date parts

A raw timestamp is hard to model; day-of-week, month, or "is weekend" turn it into features a model understands.

Binning

Group a continuous variable into ranges (age brackets, price tiers) when the effect isn't smoothly linear.

Interactions

Multiply two features when their combination matters more than either alone — e.g. region × product.

Common Mistakes

• Leakage: building a feature from information you won't have at prediction time
• Creating dozens of features with no hypothesis — noise, not signal
• Forgetting to apply the same transformation to new data

Feature Engineering in the SKARI Data Editor

In SKARI's Data Editor, the Derived Column tool builds new features from a formula — combine columns with arithmetic and functions, no code required.

• Build ratios, differences and conditionals with a visual formula editor
• Apply functions like round, log and sqrt to any column
• Every derived column is recorded in the Pipeline history for reuse

Frequently Asked Questions

More features or better features?

Better. A few well-reasoned features beat a pile of noisy ones, which only invite overfitting.

Do tree models need feature engineering?

Less than linear models, but a good ratio or date part still helps — trees can't invent a relationship the raw columns don't express.

How do I avoid leakage?

Only use information that would be available at prediction time, and build features after splitting into train and test.

Conclusion

Feature engineering is where domain knowledge meets the model. Reshape raw columns into variables that express what you actually believe drives the outcome — and guard against leakage.