What makes a model linear?

The prediction is a weighted sum of inputs: each field contributes weight times value, with no bending or interaction.

Why are linear models called interpretable?

Because every weight is a readable statement of one field's constant contribution, so the model explains each decision arithmetically.

When does linear fail?

When the true relationship curves, thresholds or depends on combinations, and the needed bend cannot be added by transforming inputs.

AI Glossary / AI and machine learning / Beginner

Linear

Definition

Linear describes a relationship that follows a straight line: double the input and you double the output's contribution, with no curves, thresholds or interactions, and by extension any model built from such straight pieces.

What it means

A linear relationship is the simplest honest one: each unit of input adds the same fixed amount to the output, wherever on the scale you stand. Ten more square meters always adds the same rent, whether the flat is small or huge.

Linear models generalize the idea to many inputs: the prediction is a weighted sum, each field contributing its weight times its value, full stop. No field's effect bends, and no two fields conspire; every contribution is independent and constant.

The simplicity buys a great deal: linear models are fast, stable, data-frugal and interpretable, since every weight is a readable statement about its field's contribution to every prediction.

The price is the ceiling: where reality curves, thresholds or interacts, a straight-line model cannot follow, and no amount of data fixes the geometry. The art is knowing when the straight answer is the right answer and when it is a confident approximation.

For a business, linear is the right default mental model: assume effects add until the evidence says they bend, because the straight model is the honest benchmark every fancier alternative must beat.

How it works

Fit the weighted sum on training data, choosing weights that best match the known outcomes.

Read the weights directly: sign gives direction, size gives strength, and together they are the model's explanation of itself.

Check the straightness assumption: plot residuals and key relationships, looking for curves and interactions the model cannot see.

Transform inputs when a known bend exists, log for proportional effects, say, so a linear fit on transformed values matches curved reality.

Benchmark any complex model against the linear baseline on held-out data before paying for complexity.

Three practical examples

A rent model where each square meter adds a fixed amount, and the residual plot stays flat across sizes: the straight story holds.
A sales model where the effect of discount clearly bends past 20 per cent, demanding a curve the linear form cannot draw.
A credit scorecard, a linear model at heart, explaining every decline as the sum of visible per-field contributions.

Think of it this way

Think of a taxi meter: so much per kilometer, the same for the first kilometer and the fiftieth, and the fare is the sum of the parts. Linear pricing is the meter; everyone understands it, and it works until the day the fare should bend, rush hour, luggage, a flight to catch.

The limit: the meter cannot see occasions. A linear model gives the same field the same say everywhere, so when the world has occasions, thresholds and combinations, the meter needs adjusting, or replacing.

Worked example

when the meter should bend

Suppose a delivery firm models cost per route linearly: fuel per kilometer, time per stop, fixed per vehicle. The figures are illustrative.

The model prices 95 per cent of routes within a few per cent, but consistently underprices the longest routes, where overtime and overnight rules kick in.

A threshold feature for routes past the overtime boundary fixes the tail, and the model stays linear in form, just better briefed.

The lesson generalizes: most curved problems are straight problems missing an occasion, and adding the occasion often beats changing the mathematics.

Illustrative case study

A fictional retailer replaced a working linear pricing model with a complex one to capture supposed subtleties, and the new model's errors concentrated exactly where the old model had been calm, the ordinary middle of the range.

Re-examination showed the subtle effects were real but tiny, and the complex model had paid for them with instability everywhere else.

The invented lesson: linear is the model of the middle. Beat it convincingly on held-out data, including the boring middle, before replacing it.

Common misconceptions and mistakes

Linear means simplistic and weak. On well-prepared business data, linear models regularly match far heavier machinery, with better explanations.
A linear model cannot handle many fields. It handles thousands happily; the limit is shape, not width.
If the data is big enough, linearity stops mattering. Volume does not bend geometry: a wrong-shape model at scale is a wrong-shape model with confidence.

Business use

Default to a linear baseline in every modeling exercise, and make alternatives prove themselves against it.

Read weight reports in reviews: they are the model's own account of what it listens to.

Look for occasions, thresholds and interactions before abandoning the linear form; transformed features often rescue it.

Keep the linear model in production as a shadow even after upgrading; divergence between the simple and complex models is an early warning worth having.

Check your understanding

Why is a linear model the right default benchmark?

Answer: Because it is fast, stable and self-explaining, so any more complex model must justify its cost by beating the straight answer on unseen data, not merely on training data.

Related terms

Learn next

AI for Business Owners covers how to judge when a simple model is enough and when complexity earns its keep: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Line spectral pairs Next: Linear belief function