What dimension is a feature vector?

However many fields the feature set has: a 40-field set produces 40-dimensional vectors.

How does similarity of vectors drive recommendations?

Neighbor methods find vectors close to yours and recommend what those neighbors liked.

Is a feature vector the same as an embedding?

An embedding is a learned feature vector; the term covers both hand-built and learned forms.

AI Glossary / AI and machine learning / Beginner

Feature vector

Definition

A feature vector is one record's values across all the model's input fields, written as an ordered list of numbers: that customer, order or event, as the model sees it.

What it means

If the feature matrix is the whole table, the feature vector is one row of it: all the field values for a single case, in the agreed column order. It is the unit the model actually consumes, one prediction at a time.

The vector is the case translated into the model's language: age 44, tenure 31 months, last contact 6 days ago becomes a string of numbers whose positions mean specific fields.

Vectors are what similarity means in practice: two customers whose vectors are close in feature space are, by the model's lights, alike, which is the basis of recommendations, segmentation and neighbor-based prediction.

The vector's health is the prediction's health: a missing value, a wrong unit or a misaligned field in the vector silently poisons one answer, and no error message appears.

For a business, the feature vector is the right unit for auditing individual decisions: pull the vector behind a surprising prediction and you can see exactly what the model was told.

Vectors also enable one of machine learning's oldest tricks: distance between vectors is a working definition of similarity, and much of recommendation, retrieval and segmentation is that one idea, dressed up.

How it works

Assemble the case's raw data and compute each feature through its defined function, in the feature set's agreed order.

Validate the assembled vector: types, ranges and completeness, before it reaches the model.

Score it: the model maps the vector to a prediction, and the vector plus prediction plus timestamp form the decision record.

Store the scored vectors for consequential decisions, so any answer can later be reconstructed and audited.

Watch live vectors against training profiles: a vector far outside the training region is a prediction being made on thin ice.

Three practical examples

A fraud score computed from a 40-number vector describing one transaction.
A recommendation made by comparing a customer's vector against vectors of similar customers.
An audit that retrieves the exact vector behind a declined application to explain the decision.

Think of it this way

Think of a patient's observation chart at the end of a hospital bed: temperature, pulse, pressure, each in its box, in the same order every time. The doctor reads the chart, not the patient, in the first instance, and the model likewise reads only the vector.

The limit: the chart has boxes only for what someone decided to measure. A vector says everything about the fields it has and nothing about the ones it lacks.

Worked example

auditing a declined application

Suppose a customer asks why their loan application was declined, and the bank retrieves the feature vector scored that day. The figures are illustrative.

The vector shows a debt-to-income value far above the applicant's real position: a data join had double-counted one liability.

The error is corrected, the application rescored and approved, and the incident becomes a case study in keeping scored vectors.

Without the stored vector, the answer would have been a shrug; with it, the answer was a fix.

Illustrative case study

A fictional retailer's recommendation engine kept suggesting barbecue equipment to a vegetarian customer, and the explanation team traced it to one vector field: a miscoded dietary flag.

One wrong number in one vector had steered months of recommendations, invisibly, because nobody ever looked at the inputs behind an odd suggestion.

The invented lesson: when an individual answer looks mad, audit the vector, not the model. The vector is where one case's truth lives or dies.

Common misconceptions and mistakes

The vector is the raw record. It is the transformed, model-ready version; the raw record is its ancestor, not its equal.
Models see whole tables at once. At prediction time they consume one vector per case; batch processing is just many vectors in sequence.
A valid vector means a correct prediction. A clean vector only guarantees the model was told the truth as defined; the model can still be wrong.

Business use

Store scored feature vectors for any decision a customer can question; they are the only faithful record of what the model knew.

Add vector-level validation to scoring: range and completeness checks catch one-off poison that aggregate monitoring misses.

Use vector explanations in complaints handling: showing the fields behind a decision defuses most disputes faster than defending the model.

When two systems score the same customer differently, compare their vectors field by field first; divergent inputs explain most divergent answers before any model blame is needed.

Include a few real vectors, anonymized, in model documentation; reviewers understand a model faster from three rows of its actual inputs than from thirty pages about its architecture.

Check your understanding

Why is the stored feature vector the best evidence for explaining one specific model decision?

Answer: Because it records exactly what the model was told about that case, so any error in the decision can be traced to the inputs, the model, or both.

Related terms

Learn next

Be Productive Using AI covers the habit of checking the inputs behind any surprising automated answer: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S2] Google, Machine Learning Glossary; [S7] Hopsworks, MLOps Dictionary; [S8] Jurafsky and Martin, Speech and Language Processing (3rd ed. draft); [S12] Aalto University, Dictionary of Machine Learning; [S24] scikit-learn documentation, Glossary of Common Terms and API Elements. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Feature template Next: Federated averaging