AI Glossary / AI and machine learning / Beginner

Feature cross

Definition

A feature cross is a new input made by combining two or more existing fields, so the model can learn patterns that only appear when the fields interact.

What it means

Single fields can miss the point. Age alone predicts little about holiday spending, and day of week alone predicts little, but young travelers on a Friday behave very differently from retirees on a Tuesday. The pattern lives in the combination.

A feature cross manufactures that combination as an explicit input: multiply two numbers, concatenate two categories into a compound category such as channel-by-region, or flag when two conditions coincide.

Linear-style models need crosses most, because they cannot discover interactions for themselves: every curved or combined relationship must be handed to them. Tree-based models find many interactions on their own, which is one reason they dominate on business tables.

The danger is explosion. Crossing ten fields with ten values each creates a hundred compound values; crossing freely creates thousands of sparse, barely-populated columns that invite overfitting.

For a business, the best crosses are hypotheses: someone suspects weekend mobile shoppers behave differently, builds the cross, and lets held-out data confirm or kill the idea.

A practical naming habit helps: compound fields such as channel__by__season read as hypotheses in the feature list, which makes reviews faster and stops crosses masquerading as raw fields.

How it works

Identify two or more fields whose joint behavior might matter, ideally from a business hypothesis rather than brute force.

Construct the combined field: a product of numbers, a compound category joining values, or an indicator for the coincidence of conditions.

Control the combinatorial growth: cross few fields, group rare compound values into an other bucket, and drop crosses with too little data.

Test each cross on held-out data: keep the ones that improve real performance, discard the rest, and record the decision.

Prefer letting interaction-finding models, trees and their ensembles, discover crosses automatically when you have no hypothesis to test.

Three practical examples

• Retail: crossing customer age band with product category to capture life-stage buying patterns.
• Logistics: crossing day of week with region to model weekend delivery demand where it actually occurs.
• Marketing: crossing device type with hour of day, catching the commuter-phone effect on email response.

Think of it this way

Think of a chef tasting two ingredients separately and then together. Lemon is sour, garlic is pungent, and lemon-garlic is a third thing entirely that neither ingredient predicts alone. A feature cross is writing lemon-garlic onto the menu.

The limit: a menu of every possible pairing is unusable. Crosses earn their place like dishes do, by being ordered, which in modeling means surviving held-out testing.

Worked example

testing a channel-season cross

Suppose a retailer's response model uses channel and season as separate fields and misses that email works unusually well in the autumn sale period. The figures are illustrative.

The team builds a compound field, channel-by-season, with twelve combined values, and adds it beside the originals.

Held-out response accuracy improves 2 per cent, and inspection shows the email-autumn cell carries the gain.

A second hypothesized cross, channel-by-region, adds nothing and is dropped, which the team counts as a cheap and useful answer.

Illustrative case study

A fictional bank crossed every pair of its forty applicant fields, producing eight hundred compound columns, and its model's training accuracy soared while real accuracy quietly fell.

The crosses had memorized tiny intersections: left-handed plumbers from one town, patterns with three examples each.

Pruning to six hypothesized crosses recovered the gain without the noise. The invented lesson: a cross is a claim that two fields interact. Make claims you can defend, not every claim arithmetic allows.

Common misconceptions and mistakes

• Feature crosses are only for statisticians. A compound field built in a spreadsheet is a cross; the concept is simple even when the libraries are formal.
• More crosses mean more power. Each cross multiplies sparsity and overfitting risk; a few hypothesized crosses beat a blanket of them.
• Trees make crosses unnecessary everywhere. Trees find interactions in numeric splits, but explicit crosses still help linear models and can sharpen even tree models on categorical interactions.

Business use

Source your crosses from people who know the business: the best interactions are suspected before they are measured.

Set a budget for compound fields and enforce it, because crossing is the fastest way to turn forty tidy columns into four thousand problems.

Review surviving crosses with domain owners; an interaction the business cannot explain is an interaction it cannot defend.

When a cross survives testing, translate it back into plain language for the decision makers: the value of the finding is the rule it suggests, not the column it created.

People also ask

How is a feature cross different from a polynomial term?

A polynomial term combines a field with itself; a cross combines different fields, though both let simple models represent curved or joint relationships.

Do deep networks need crosses?

Rarely; they learn interactions internally. Crosses matter most for linear models and small-data problems.

How do I limit cross explosion?

Cross few fields, bucket rare combinations, and keep only crosses that survive held-out testing.

Check your understanding

Why do linear models benefit most from explicit feature crosses?

Answer: Because a linear model can only combine fields additively, so any interaction must be supplied as a ready-made combined input; flexible models can discover interactions themselves.

Related terms

• Feature engineering
• Feature interactions
• Decision tree
• Sparsity
• polynomial features
• model complexity

Learn next

AI for Business Owners explains how to turn staff hypotheses about customer behavior into inputs a model can test: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

More in AI and machine learning

• Factor graph - A factor graph represents how a function of several variables can be written as a product of smaller functions called factors.
• Farthest-first traversal - Farthest-first traversal orders points by repeatedly selecting the point whose distance to its nearest previously selected point is greatest.
• Feature - A feature is one measurable input field a model learns from, such as a customer's age, an order's total or the hour a call arrived.
• Feature cutoff - A feature cutoff is a threshold that drops feature values below a set frequency or importance, keeping the input space small enough to be useful.
• Feature extraction - Feature extraction is deriving informative measurements from raw, unstructured material, such as turning a photo into edges and textures or a document into...
• Feature extractor - A feature extractor is the component that converts raw input into feature values: the fixed recipe, pipeline stage or trained network that stands between...

Sources

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

Previous: Feature Next: Feature cutoff