AI Glossary / AI and machine learning / Intermediate
Feature interactions
Definition
Feature interactions are the ways two or more input fields affect the outcome jointly, so that the effect of one depends on the value of another.
What it means
Most fields do not act alone. Discount depth matters more for price-sensitive segments; rain affects weekend sales differently from weekday sales. The outcome depends on the combination, and that dependence is an interaction.
Interactions are where simple models go blind. A linear model adds each field's effect independently and cannot represent depends-on, so someone must hand it the interaction as a built feature, or it will average the effect away.
Flexible models, trees, forests, deep networks, discover interactions internally, but they need enough data to find them, and small or noisy data hides interactions from even the best model.
Interactions also carry the business story: which combinations drive the outcome is often more actionable than which single fields do, because interventions happen at combinations, a campaign aimed at a segment in a season.
For a business, asking what interacts with what is a high-value habit: it surfaces the compound conditions where policy, pricing and targeting should differ.
How it works
Suspect interactions from domain knowledge first: which pairs of fields plausibly modify each other's effect.
Test candidates explicitly: build the crossed feature, or fit models with and without the interaction and compare held-out results.
In tree models, read interactions from the structure: a field splitting differently beneath another field's branch is an interaction made visible.
Quantify with interaction-strength measures or partial-dependence plots, which show how one field's effect changes across another's values.
Act on confirmed interactions: segment the policy, price or message by the combination, and monitor whether the interaction persists as conditions change.
Keep a record of interactions tested and rejected, not only those kept; the negative results stop future teams from re-testing the same dead ends every year.
Three practical examples
- Insurance: age interacts with vehicle type, so young drivers of powerful cars price differently from what either field alone suggests.
- Retail: promotion depth interacts with store affluence, so the same discount pulls different crowds.
- SaaS: onboarding completion interacts with company size, so activation campaigns differ by both.
Think of it this way
Think of mixing medicine. Each drug has its effect, and the interaction is what happens when they meet, sometimes stronger, sometimes dangerous, never predictable from the labels alone. Feature interactions are the same chemistry in data.
The limit: the body is complicated and so is the market. Most pairs do not interact, which is why hypothesis-led testing beats checking every possible pair blindly.
Worked example
finding a pricing interaction
Suppose a software firm models deal size against discount and finds discount's effect surprisingly weak on average. The figures are illustrative.
A region split reveals the interaction: discounts strongly lift deals in new regions and barely move mature ones, and the average hid both.
The firm builds region-type by discount-level cells into pricing guidance, and approval workflows route deep discounts in mature regions to extra review.
Average discount depth falls two points with no volume loss, because the blunt instrument became a targeted one.
Illustrative case study
A fictional lender's scorecard treated income and loan purpose independently for years until an analyst plotted them together and found that high income reversed the risk of one loan purpose rather than merely reducing it.
The interaction had been invisible in the additive scorecard and obvious in the cross-tab.
The invented lesson: additivity is an assumption, not a law. Plot fields against each other before believing either's average effect.
The invented follow-up: the lender added interaction plots to its annual model review, and the ritual has since caught two pricing rules whose effects had quietly flipped sign as the portfolio mix changed.
Common misconceptions and mistakes
- If both fields are significant, their interaction must be. Significance of the parts says nothing about the combination; interactions must be tested on their own.
- Trees find all interactions automatically. They find interactions the data supports; sparse regions and rare combinations can hide real ones from any model.
- Interactions are statistical trivia. They are often the most actionable finding in the whole model, because decisions are made at combinations.
Business use
Ask in every model review: which interactions were tested, and which were assumed away by the model form.
Beware average effects in pricing and policy; an average can conceal an interaction that reverses the sign for a whole segment.
Give analysts tooling for interaction plots, because the finding that changes policy usually arrives as a picture, not a coefficient.
People also ask
How do you detect an interaction?
By comparing a model with the interaction term against one without on held-out data, or by plotting one field's effect across the other's values.
What is a three-way interaction?
A dependence involving three fields at once; real but hard to estimate and harder to explain, so test two-way ones first.
Do interactions imply causation?
No more than any correlation does; they mark combinations worth investigating, not proof of mechanism.
Are interactions the same as correlations?
No; correlation is between two inputs, interaction is about their joint effect on the outcome.
Check your understanding
Why can a model that includes both fields still miss their interaction?
Answer: Because additive models combine each field's effect independently, and unless the interaction is supplied or discoverable, the depends-on relationship is averaged away.
Related terms
Learn next
AI for Business Owners explains how to read past averages to the combinations where decisions actually change: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Feature Freshness - Feature freshness is how up to date a model's input values are: the gap between what the data says and what is true right now.
- Feature Function - A feature function is the reusable piece of logic that computes one feature from raw data: given the inputs, here is exactly how this field gets its value.
- Feature hashing - Feature hashing is a trick for turning huge or open-ended sets of values into a fixed number of numeric columns, by passing each value through a hash...
- Feature learning - Feature learning is letting a model discover the useful input representations itself from raw data, instead of a person designing the features by hand.
- Feature map - A feature map is the output produced when a learned detector is applied across an input, such as the grid of responses an edge detector leaves at every...
- Feature matrix - A feature matrix is the table a model trains on: one row per example, one column per feature, with each cell holding that example's value for that field.
Sources
[S8] Jurafsky and Martin, Speech and Language Processing (3rd ed. draft). Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Feature hashing Next: Feature learning