What does Multi-task Lasso share across tasks?

The field list: a field is kept or evicted for the whole family, while each task keeps its own coefficients.

How does it differ from Multi-task Elastic-Net?

Lasso uses the pure L1-style eviction; Elastic-Net mixes in L2, which keeps groups of correlated fields together.

When is it a bad idea?

When the tasks do not genuinely share drivers, because shared eviction then exports one task's noise to the others.

Data-thin tasks in a real family, which borrow the family's evidence for field selection while keeping local coefficients.

AI Glossary / AI and machine learning / Advanced

Multi-task Lasso

Definition

Multi-task Lasso is a regression method that learns several related tasks together under a shared sparsity penalty: a field is kept or evicted for the whole family at once, with each task keeping its own coefficients.

What it means

The method sharpens the multi-task idea. Instead of pooling data or fitting separate models, it fits all tasks jointly and lets the lasso penalty select fields at the family level: in for everyone or out for everyone.

The shared eviction is the point. A field that looks weak in one task's noisy data but helps across the family survives, and a field that helps only one task by accident is evicted, so the family's combined evidence steadies the selection.

The result is one field list with per-task coefficients: a common language for what matters, plus local numbers for how much, which is a natural fit for reporting to regional or product owners.

The assumption underneath is relatedness: the tasks must genuinely share drivers, or the shared eviction exports one task's noise to the whole family.

For a business, Multi-task Lasso answers the recurring question of whether to build one model or many: its answer is both, one field list, many coefficient sets, with validation deciding whether the family is real.

The penalty operates on groups of coefficients, one group per field across tasks, which is what makes the keep-or-evict decision a family decision.

How it works

Verify the tasks share drivers, by comparing separately fitted models for overlap and by domain review.

Standardize fields within each task so the penalty compares evidence rather than units.

Fit jointly across tasks with the shared penalty tuned on held-out error, checking stability under resampling.

Read the output as two products: the family field list, and the per-task coefficient tables on that list.

Benchmark against per-task models and a pooled model, and adopt only if the joint model beats both on held-out data.

Three practical examples

A bank scoring default risk per product line with one shared field list, where thin products borrow evidence from rich ones.
A retailer forecasting per-store demand with store-specific coefficients on a shared weather, price and calendar field list.
A clinic group predicting readmission per site, sharing a clinical field list while letting each site weight it locally.

Think of it this way

Think of a hiring panel choosing one shortlist for several branches: every branch interviews from the same list but makes its own offer, so weak candidates are filtered once, collectively, and each branch still decides for itself.

Worked example

the thin product that borrowed strength

Suppose an insurer prices five product lines, and the newest line has six months of data against five years elsewhere. The figures are illustrative.

A standalone model for the new line picks noisy, unstable fields, changing with every month's data and eroding trust in the pricing.

Multi-task Lasso fits all five lines together, and the new line inherits the family's stable field list while keeping its own coefficients, steadying its model immediately.

The gain is real but conditional: it rests on the lines genuinely sharing drivers, which the team verifies by checking that the shared list matches actuarial judgment.

Illustrative case study

A fictional retailer applied Multi-task Lasso across grocery and pharmacy categories, assuming one retail logic, and the shared penalty evicted fields that mattered enormously to pharmacy but little to grocery.

Pharmacy forecasts degraded while grocery improved, and the average masked the split until the pharmacy team ran its own numbers.

The invented lesson: shared eviction is a strong claim of family. When the family is false, the penalty does not average the tasks; it sacrifices the minority one.

Common misconceptions and mistakes

It pools the data into one model. It fits jointly but keeps per-task coefficients; only the field list is shared.
Thin tasks always benefit. They benefit when related; a thin stranger task is harmed, not helped, by the family.
The shared list is permanent. Field selection is as unstable as any data-driven choice and needs resampling checks and periodic review.

Business use

Adopt Multi-task Lasso when several genuinely related tasks share candidate fields and at least one task is data-thin.

Test relatedness first, and give a demonstrably different sub-family its own model rather than forcing membership.

Report the shared field list and per-task coefficients separately; they answer different questions for different audiences.

Check the selection's stability under resampling before treating the field list as settled fact.

Check your understanding

How does Multi-task Lasso help a data-thin task, and what must be true for the help to be real?

Answer: The thin task borrows the family's combined evidence for choosing fields, steadying its model, but only if the tasks genuinely share drivers; with a false family, shared eviction imports the other tasks' noise and sacrifices the minority task's real structure.

Related terms

Learn next

AI for Business Owners helps you decide when models should share structure and when they should stand alone: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S24] scikit-learn documentation, Glossary of Common Terms and API Elements; [S13] Minka, Statistical Learning/Pattern Recognition Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Multi-task Elastic-Net Next: Multiagent planning problem