AI Glossary / AI and machine learning / Advanced
Elastic-Net
Definition
Elastic-Net is a regression method that combines the two classic penalty types, L1 and L2, so it selects variables like the first and stabilizes estimates like the second.
What it means
Linear regression has two famous repairs for unreliable coefficients. Ridge regression shrinks all coefficients towards zero, which stabilizes them but keeps every field in the model. Lasso drives some coefficients exactly to zero, which selects fields but behaves erratically when fields are correlated, picking one of a group of similar fields at random.
Elastic-Net blends the two penalties with a mixing dial. The L1 part still zeroes out useless fields; the L2 part stops the selection from being arbitrary among correlated fields, tending to keep or drop similar fields together.
The result suits exactly the situation business data is usually in: many candidate fields, thick with correlation, some genuinely useless.
Two settings must be tuned, the overall penalty strength and the L1-versus-L2 mix, which is one dial more than either parent method and one more place for careless defaults.
For a business, Elastic-Net is a workhorse for interpretable prediction: a pricing or demand model whose surviving coefficients can be read and defended in a meeting.
How it works
Start with standardized input fields, because penalties treat fields differently if their scales differ.
Set the penalty strength, which controls how hard coefficients are pushed towards zero, and the mixing ratio between the L1 and L2 penalties.
Fit the model by minimizing prediction error plus the blended penalty, which shrinks weak coefficients to exactly zero and moderates the rest.
Choose the two settings by cross-validation, testing a grid of strengths and mixes against held-out folds.
Read the surviving coefficients as the model's variable selection, checking that the kept fields make domain sense before trusting the predictions.
Inspect the coefficient paths as the penalty strength varies: fields entering and leaving the model as the dial turns tell you which selections are stable and which live at the margin.
Three practical examples
- Pricing models: selecting the few deal attributes that genuinely move willingness to pay from dozens of correlated candidates.
- Demand forecasting: keeping whole families of related promotion variables together rather than arbitrarily choosing one.
- Credit scoring: a sparse, stable scorecard when applicant fields overlap heavily.
Think of it this way
Think of two editors with different tempers. One deletes whole paragraphs without hesitation; the other trims every sentence a little. Elastic-Net hires both: whole useless sections go, and the surviving text is evenly tightened.
The limit: two editors need managing. The mix between them is a judgment, and the wrong mix inherits the worst of either temper.
Worked example
tuning a pricing model
Suppose a manufacturer models deal margin against 80 candidate fields, many measuring overlapping aspects of deal size and customer history. The figures are illustrative.
Lasso alone keeps 14 fields but drops three that the sales team insists matter, choosing arbitrary single representatives from correlated groups.
Elastic-Net with a 70-30 L1-L2 mix keeps 19 fields, including the whole correlated deal-size family, and cross-validated error improves by 4 per cent.
The final coefficient table is short enough to review in one meeting, and the sales team recognizes every field on it.
Illustrative case study
A fictional insurer built a risk scorecard with lasso and found its selected fields changed every time the training data was refreshed, because correlated fields were being chosen by coin toss.
Switching to Elastic-Net stabilized the selection across refreshes, which mattered more than the tiny accuracy difference, because regulators asked why fields came and went.
The invented lesson: when correlated fields compete, pure selection is a lottery. Stabilize first, then select.
The invented follow-up: the insurer now reports selection stability, how often each field survives resampling, alongside accuracy, because a scorecard whose membership changes quarterly cannot be explained to anyone twice the same way.
Common misconceptions and mistakes
- It always beats lasso and ridge. It usually helps when fields are correlated and numerous; on clean, nearly independent fields the parents can do as well.
- The mix is fifty-fifty by nature. The mix is a tuned setting, and the best value is often far from even.
- Zero coefficients mean irrelevant fields. They mean the penalty found them unhelpful at this strength; a field can be zeroed and still matter at a different setting.
Business use
Use Elastic-Net as the default interpretable baseline when fields are many and correlated; it rarely embarrasses and often selects sensibly.
Report the mixing and strength settings alongside any coefficient table, because the table is meaningless without them.
Check selection stability across data refreshes; a model whose chosen fields churn is a model whose explanations churn.
People also ask
What do L1 and L2 mean?
Two penalty shapes: L1 penalizes the absolute size of coefficients and can zero them; L2 penalizes their squared size and only shrinks them.
Is Elastic-Net only for linear models?
The penalty blend generalizes, but the named method refers to the linear case, where it is standard.
How is the mix chosen?
By cross-validation over a grid of values, like any other tuning setting.
Check your understanding
Why does Elastic-Net tend to keep groups of correlated fields together when lasso does not?
Answer: Because its L2 component penalizes arbitrary concentration on one of several similar fields, so the selection spreads across the group instead of picking a random single representative.
Related terms
Learn next
AI for Business Owners explains how to read a coefficient table and challenge a model's choice of variables: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Eager learning - Eager learning builds a general predictive model from training examples before new prediction requests arrive.
- Early stopping - Early stopping is ending a model's training at the point where it performs best on fresh data, before it starts memorizing the training set.
- Einsum notation - Einsum notation is a compact way of writing operations on multi-dimensional arrays by naming their axes, so sums, products, transposes and contractions are...
- ELMo - ELMo, short for Embeddings from Language Models, is a method for producing word representations that change with the surrounding text.
- Embodied agent - An embodied agent perceives and acts through a body situated in an environment.
- Emergent algorithm - An emergent algorithm produces a larger pattern of behavior through interactions among simpler components following local rules.
Sources
[S24] scikit-learn documentation, Glossary of Common Terms and API Elements. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Elastic map Next: Electronic image stabilization