What does lasso stand for?

Least Absolute Shrinkage and Selection Operator: shrinkage of coefficients and selection of variables in one penalty.

How is it different from ridge?

Ridge, the L2 penalty, shrinks all weights but keeps all fields; lasso drives weak fields to exactly zero.

What is the LARS Lasso?

The lasso fitted by the LARS algorithm, which computes the whole path of models efficiently as the penalty relaxes.

AI Glossary / AI and machine learning / Intermediate

Lasso

Definition

Lasso is regression with an L1 penalty, famous for driving the weights of unneeded fields to exactly zero, so fitting the model and selecting its fields happen in one step.

What it means

The name is an acronym, Least Absolute Shrinkage and Selection Operator, and the two halves are the whole story: it shrinks coefficients and it selects variables, and it does both by charging a tax on the absolute size of every weight.

The tax has a distinctive geometry: where gentler penalties shrink weights toward zero asymptotically, the lasso's penalty makes zero itself attractive, so weak fields are evicted outright rather than merely quietened.

The result is a sparse model and a survivor list: out of hundreds of candidate fields, the lasso keeps the ones that earn their rent, which is often the single most valuable output for a business trying to understand its own drivers.

The penalty strength is the dial: higher values evict more aggressively, and the right setting is found by validation, with the full path of models, from empty to full, available from efficient algorithms like LARS.

For a business, the lasso is the standard first answer to which of these hundreds of fields matter: fast, interpretable, honest about uncertainty through validation, and humble enough to say these few, not these many.

It sits inside a family: the L1 penalty in other models earns the same lasso behavior, and hybrids like the elastic net blend its evictions with ridge's stability for correlated fields.

How it works

Standardize candidate fields so the penalty judges them on merit, not on unit size.

Fit across a range of penalty strengths, or compute the full path, and score each point on validation data.

Choose the strength where held-out error bottoms out, and read off the survivor fields and their coefficients.

Inspect evictions of fields you believed mattered: they usually signal a stronger correlated rival rather than a useless field.

Refit at the chosen strength, validate once more on sealed data, and document the survivor list.

Three practical examples

A pricing model distilled from 120 candidate drivers to 19 survivors, each reviewed and named by the business.
A text model keeping 900 of 50,000 word features, with the survivor list doubling as a vocabulary of what predicts.
A regional demand screen where lasso survivors differ by region, revealing genuinely different local drivers.

Think of it this way

Think of a football squad chosen under a hard salary cap. Every player must justify the wage from current form, and those who cannot are released, not benched. The lasso is the cap: it forces the squad to its earning members.

The limit: the cap judges players against their teammates. Two similar strikers split the evidence and one is released, so a surprising release usually means look at who got the jersey, not the released player was never good.

Worked example

nineteen survivors

Suppose an insurer fits a lapse model over 140 candidate fields with the lasso. The figures are illustrative.

Nineteen fields survive at the validated penalty, and the list matches actuarial intuition except for two surprises.

One surprise proves to be leakage and is removed; the other, a service-contact pattern, is genuine and becomes a retention trigger.

The nineteen-field model validates within a point of the full-field version and costs a fraction of the monitoring effort.

Illustrative case study

A fictional consultancy ran the lasso once at an arbitrary penalty and delivered the survivors as the drivers of churn, with no validation curve and no resampling.

Two of the eleven survivors vanished under mild resampling, the client's strategy had been pointed at noise, and the rework consumed the fee.

The invented lesson: the lasso's list is only as solid as the process behind the penalty. Validate the strength and test stability before naming drivers.

Common misconceptions and mistakes

Lasso finds the true drivers. It finds fields that predict well together under this penalty; truth about causes needs experiments, not evictions.
Evicted fields are useless. They lost to correlated rivals or the penalty's austerity; many would predict well in other company.
It is outdated. It remains the standard sparse linear method and a permanent fixture of the first week on any wide dataset.

Business use

Default to a lasso pass on any wide candidate list, and present the survivor list with its validation curve.

Check survivor stability under resampling before naming business drivers; unstable survivors are hypotheses, not findings.

Standardize first and say so: an unstandardised lasso ranks units, not importance.

Pair the survivor list with domain review before publishing: two lines of expert context turn a coefficient table into a finding the business will act on.

Check your understanding

Why does the lasso produce a survivor list rather than just smaller coefficients?

Answer: Because its absolute-value penalty makes exactly zero the cheapest value for weak weights, so fields that do not earn their rent are evicted, not merely discounted.

Related terms

Learn next

AI for Business Owners covers how to use a survivor list as a starting hypothesis, not a final truth: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S24] scikit-learn documentation, Glossary of Common Terms and API Elements. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: LARS Lasso Next: Late Binding