Is an instance the same as a row?

In tabular data, yes: one row, one instance. The term generalizes to one image, one document, one sequence, whatever the unit of prediction is.

What is a labeled instance?

An instance carrying the known answer, used in training and evaluation, as opposed to an unlabeled one awaiting prediction.

Why store scored instances?

So any individual decision can be reconstructed, explained and corrected, which averages alone can never provide.

What is an instance in image data?

One image, or one labeled object within it, depending on the task; the principle is the same: the single unit the model learns from or scores.

AI Glossary / AI and machine learning / Beginner

Instance

Definition

An instance is a single example in a dataset: one row, one record, one case, the unit a model learns from at training time and scores at prediction time.

What it means

Machine learning talks about datasets in the abstract, but everything concrete happens to instances: this customer, this transaction, this image. The dataset is just the collection; the instance is the atom.

Each instance arrives as a bundle of features plus, in supervised work, a label: the input fields and the known answer. The model learns by adjusting itself across many instances and predicts by consuming one instance at a time.

Vocabulary drifts here: row, record, example, sample, observation, case and instance all mean the same thing in most conversations, and the interchangeable terms confuse newcomers far more than the idea itself.

The instance is also where accountability lives. Aggregate metrics describe the dataset, but a complaint, an audit or an explanation always comes down to one instance: what did the model see for this customer, and what did it say?

For a business, keeping instances retrievable is the practical lesson: store what was scored for each case, so that any individual decision can be reopened, explained and, when wrong, corrected.

The level of granularity is a design choice: whether one instance means one transaction, one customer-day or one customer is decided before modeling, and that decision shapes everything the model can learn and every question it can later answer.

How it works

Structure data so each instance is one row with a stable identifier, its feature values and, where applicable, its label.

Keep instances consistent: the same fields, in the same units, with the same definitions across training and scoring.

Store scored instances for consequential decisions, inputs and outputs together, so individual cases stay auditable.

Handle broken instances deliberately: missing values, malformed fields and duplicates each get a rule, not a shrug.

Respect instances as people where they are people: access, retention and correction obligations attach to the individual record.

Three practical examples

A churn model scoring one instance, this subscriber's row of features, to produce this subscriber's risk score.
An auditor retrieving the exact instance behind a declined claim, features and output together, to answer the customer's complaint.
A data quality report counting instances with missing income values, because row-level defects hide inside healthy averages.

Think of it this way

Think of a hospital ward. The ward's statistics, average temperature, average stay, describe the crowd, but medicine happens to one chart at a time, and the chart is the instance: this patient's numbers, this patient's record, this patient's prescription.

The limit: charts get copied, muddled and merged. Duplicate and inconsistent instances are the everyday sickness of real datasets, and they teach models the world as it was recorded, not as it was.

Worked example

one row under the microscope

Suppose a customer disputes a declined credit line and the bank pulls the instance. The figures are illustrative.

The stored record shows the exact features scored: income, tenure, balances, and the resulting decision, all from the day in question.

One field stands out, a balance ten times reality from a feed error, the instance is corrected, rescored and approved.

The aggregate model never changed; the fix was instance-level, and it was possible only because the row was kept.

Illustrative case study

A fictional retailer merged two customer databases and let duplicate instances, the same person twice with slight spelling variants, flow straight into model training.

The model quietly learned that certain customers buy everything twice, forecasts inflated, and nobody caught it until the stock arrived.

The invented lesson: datasets are made of instances, and instances are made by fallible pipes. Count the duplicates before trusting the averages.

Common misconceptions and mistakes

An instance is a technical detail for engineers. It is the unit of every decision, explanation and dispute; business accountability happens at instance level or nowhere.
Averages describe what the model does. Averages describe tendencies; the customer experiences one instance, and so does the regulator.
Fixing instances is janitorial work. Instance quality sets the ceiling on model quality; the cleaning is the modeling, on most real projects.

Business use

Assign every instance a stable identifier and keep the scored record for any decision a customer can question.

Measure instance-level defects, missing, malformed, duplicated, as standing data quality metrics.

Build complaint and audit workflows that start from the single instance, not from the aggregate dashboard.

Check your understanding

Why does accountability for automated decisions live at instance level?

Answer: Because a dispute, audit or explanation always concerns one specific case, and only the stored instance, its inputs and output, can show what the model actually decided and from what.

Related terms

Learn next

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

Sources

[S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Insider attack Next: Instance-based learning