Interaction Term Construction¶
Feature synthesis — instantiates Task-Legible Feature Construction
Manufactures combined features — products, ratios, or conditionals of two or more raw inputs — to expose a joint effect that neither input reveals alone.
Some signals live between variables, not in any one of them. Interaction Term Construction manufactures new features by combining two or more raw inputs — a product, a ratio, a conditional, a difference-in-context — to make a joint effect visible to a consumer that would otherwise miss it. Its defining move is combinatorial and hypothesis-driven: it starts from a stated guess that "the effect of A depends on B" and builds the term that lets a model see that dependence. A linear model, in particular, cannot represent "square footage matters more downtown than in the exurbs" unless someone hands it sqft × downtown. This mechanism is where that term gets built. Unlike a trusted domain formula, an interaction is a candidate — a conjecture about a relationship, constructed to be tested, not transcribed from established practice.
Example¶
A real-estate analytics team is modeling home sale prices for a linear pricing model. Square footage helps; location helps; but the model systematically underprices large homes in premium neighborhoods and overprices large homes in cheap ones. The raw inventory tells them why the model is stuck: each input is present, but their joint effect is invisible to an additive model. The team writes the hypothesis explicitly — "the price-per-square-foot premium is not constant; it scales with neighborhood desirability" — and constructs the interaction square_footage × neighborhood_price_index.
Adding that single term lets the model bend: now an extra 500 square feet is worth far more in a top-decile neighborhood than a bottom-decile one, exactly matching what appraisers know. They also try bedrooms × bathrooms (a plausible layout-quality proxy) but it barely moves anything, so it is left as a candidate rather than promoted. The constructed terms are drawn from the raw inventory's available fields and each is tied to a written rationale, so a reviewer can see why each cross-product was built — not just that it happened to help.
How it works¶
The mechanism turns a relationship hypothesis into a concrete combined column:
- State the interaction hypothesis. Name the two-or-more inputs and the joint pattern conjectured — "A's effect depends on B." This is what separates a principled interaction from blind cross-multiplication.
- Confirm the raw inputs. Check the inventory that the constituent fields exist at the needed grain, units, and decision-time availability before combining them.
- Choose the combination form. Product (multiplicative modulation), ratio (relative scaling), difference (contrast), or conditional/segmented term. The form encodes the shape of the hypothesized dependence.
- Emit as a testable candidate. Add the term to the catalog as a candidate for downstream evaluation — its worth is decided elsewhere, by an actual value test.
The distinguishing act is constructing a joint term from a stated dependence hypothesis, aimed squarely at consumers (linear and additive models) that cannot infer interactions themselves.
Tuning parameters¶
- Interaction order — pairwise, three-way, or higher. Higher orders capture richer dependence but explode combinatorially and overfit fast.
- Combination form — product vs. ratio vs. conditional. Each imposes a different functional shape; the wrong form can hide a real interaction that another would reveal.
- Hypothesis-driven vs. exhaustive — build only theorized crosses, or generate all pairs and filter. Exhaustive search finds surprises but invites spurious terms and multiple-comparison risk.
- Standardization of constituents — whether inputs are centered before multiplying. Centering changes what the main effects mean and can stabilize the term's interpretation.
When it helps, and when it misleads¶
Its strength is unlocking signal that additive models structurally cannot see; a single well-chosen interaction can fix a whole class of systematic errors, and because it is built from named inputs it stays interpretable. It is the direct remedy for "the target signal is present only after combining inputs."
Its failure mode is combinatorial overfitting: generate every pairwise product on a modest dataset and some will correlate with the target by chance, inflating apparent performance and fragmenting under drift.[n1] The classic misuse is atheoretical cross-multiplication — building thousands of interactions and keeping whatever scores, which manufactures spurious structure and multiplies false discoveries. There is also a subtler trap: an interaction can encode a proxy relationship (e.g., zip × age) that no single input reveals. The guarding discipline is to prefer hypothesis-driven terms with a written rationale, to keep interaction order low, and to route every candidate through an independent value test before it is trusted — construction proposes, it does not confirm.
How it implements the components¶
candidate_transformation_catalog— it contributes the interaction/cross-term family of admissible transformations to the catalog.target_regularity_hypothesis— each term begins as a stated conjecture about a joint dependence the representation should expose.raw_observation_inventory— it draws on the inventory to confirm the constituent inputs exist at the right grain and availability before combining them.
This construction does NOT implement domain_expert_review_gate or feature_semantics_record for a pre-trusted formula — transcribing a ratio the field already endorses is Domain-Derived Feature Template, its nearest twin; the difference is that an interaction term is a conjecture built to be tested, while a domain template is a formula already vouched for by expert review.
Related¶
- Instantiates: Task-Legible Feature Construction — supplies the joint/cross-term entries in the transformation catalog.
- Sibling mechanisms: Domain-Derived Feature Template · Categorical Encoding Scheme · Normalization & Scaling Pipeline · Lag & Window Feature Extraction · Leakage Scan · Feature Ablation Comparison · Feature Importance & Stability Dashboard · Feature Store Versioning
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Interaction Term Construction operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it manufactures combined features — products, ratios, or conditionals of two or more raw inputs — to expose a joint effect that neither input reveals alone
Independent corroboration: The frozen evidence defines Interaction Term Construction as 'Manufactures combined features — products, ratios, or conditionals of two or more raw inputs — to expose a joint effect that neither input reveals alone', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Product and conditional terms that represent non-additive effects are canonical regression and experimental-design constructs.
Related originating lineages:
- Data Science & Analytics — Feature engineering materially broadens interaction construction to predictive models and automated candidate generation.
Review resolution: Both independent reviews place the primary lineage in statistics_experimental_design. The queued differences (origin_mode_disagreement, domain_reach_disagreement) concern secondary metadata rather than primary provenance. The final retains data_science only where a reviewer supplied a formative-lineage rationale; this does not convert downstream applicability into origin. origin_mode=convergent because the reviewers document independently established or materially co-developing traditions. domain_reach=multi_domain records application breadth separately from provenance.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] An interaction effect exists when the effect of one variable on an outcome depends on the level of another; it cannot be captured by a purely additive model without an explicit product or conditional term. Constructing many such terms atheoretically risks spurious fits, which is why interactions are treated as candidates to be validated, not accepted on offline lift alone. ↩