Definition
Features are the columns or signals a model sees — age, word counts, embedding vectors, or rolling averages. Good features make simple models strong; bad features waste complex ones.
Deep learning learns features automatically, but tabular ML and classical NLP still benefit heavily from human-designed features.
In simple terms
Feature engineering is choosing what to measure before guessing the outcome — like a doctor ordering the right blood tests instead of guessing from appearance alone.
Where you see it
- Credit scoring uses payment history and utilization ratios as features.
- Spam filters once relied on keyword counts and sender reputation.
- Time-series models use lag and seasonality features.
How it works
1.Understand the task
What signal predicts the target?
2.Transform raw data
Encode categories, scale numbers, extract text stats.
3.Validate
Check correlation, leakage, and stability over time.
Why it matters
- Strong features often beat fancier algorithms on structured business problems.
Often confused
Neural networks made feature engineering obsolete.
It evolved — embeddings replace hand counts for text, but domain knowledge still guides what data to feed models.