Must every output use the same class names?

No. Different targets can have different category sets.

Is this identical to multilabel classification?

No. The usual multilabel representation uses binary presence decisions, while these targets can each be multiclass.

Does a multi-output wrapper necessarily model target dependencies?

No. A basic wrapper can fit a separate classifier for each target.

AI Glossary / MLOps and deployment / Advanced

Multi-class multi-output

Definition

Multi-class multi-output classification predicts several categorical target variables for each input, with targets capable of taking more than two class values. For example, one record might require both a department category and a priority category. Each output has its own set of possible classes. The task differs from ordinary single-output multiclass classification and from the usual multilabel setting of multiple binary label-presence decisions.

What it means

One input can require more than one answer. The answers need not share the same categories, and success on one does not imply success on another. A model might correctly identify which department should receive a request while assigning the wrong urgency. Describing the task as multi-output makes this structure explicit. It does not require one particular architecture: separate classifiers, a shared model with multiple output heads or other structured approaches can be considered depending on the relationships and data.

How it works

Training examples pair each input with a vector of categorical targets. A simple approach fits one classifier per target, while other methods can share representations or model dependencies. The target encoding must preserve which column represents which task and which class identifiers belong to it. Evaluation can report performance for each output and a stricter measure requiring all outputs to be correct for a record. These measures answer different questions. Missing labels and invalid combinations require explicit handling rather than being silently treated as ordinary observed classes.

Three practical examples

A service system predicts both request type and priority, with several possible categories for each.
A product-image system predicts an item's category and condition class from the same image.
A document workflow predicts document type and destination team, then evaluates each output separately as well as their combined correctness.

Think of it this way

Imagine filling out several categorical fields on a form after reading one description. Each field asks a different question and offers its own answer choices. Completing the form involves multiple classification decisions, even though the input is shared. Some answers may be related, but the relationship should be learned or specified rather than assumed. A completed form can be partly correct, which explains why per-field accuracy and the percentage of entirely correct forms are distinct measures.

Worked example

Use an original test set of four requests and two outputs: department and priority. Suppose department predictions are correct for requests 1, 2 and 3, giving 3/4 = 75 percent accuracy. Priority predictions are correct for requests 1, 2 and 4, also giving 75 percent. Both outputs are correct only for requests 1 and 2, so the all-outputs-correct rate is 2/4 = 50 percent. Across the eight individual output decisions, six are correct, again 75 percent. These figures demonstrate how a useful-looking average can coexist with fewer completely correct records. They are stipulated results, not evidence of an actual model's performance or class balance.

Illustrative case study

A fictional support team builds a model that predicts request category and priority. Its first report combines all output decisions into one score, hiding that priority mistakes have a larger operational effect. The team adds separate confusion analyses and an all-outputs-correct measure, then reviews cases where predictions form implausible combinations. It compares an independent-classifier baseline with a shared-representation approach using held-out examples. The scenario illustrates clearer task definition and evaluation, without claiming that sharing representations always improves accuracy or that every combination of output labels should be allowed.

Common misconceptions and mistakes

Multi-class multi-output is not simply selecting several labels from one binary indicator list, although both involve multiple decisions. It also does not mean predicting several continuous numbers, which is multi-output regression. Another mistake is assuming the outputs are statistically independent because they occupy separate columns. A simple independent-classifier strategy may ignore useful relationships. Conversely, forcing every combination into a single large class can create a sparse and unwieldy label space. Finally, one aggregate score can conceal a weak output or an important minority class.

Business use

The concept is useful when one business record requires several categorical decisions. Define each target, its allowed classes and the operational consequences of mistakes. Establish which combinations are valid and whether missing labels are common. Evaluate each output and any joint requirement on representative held-out data, including relevant subgroups and rare classes. Compare a straightforward baseline before adding complexity. A useful deployment decision depends on the actual workflow: partial correctness may be sufficient for one use but unacceptable where all predicted fields jointly determine an action.

Check your understanding

What was the department accuracy in the example?

Answer: 75 percent.

Why was the all-outputs-correct rate only 50 percent?

Answer: Only two requests had both predictions right.

What did the 75-percent average across decisions fail to show?

Answer: How often a complete pair of answers was correct for the same request.

Related terms

Multiclass classification
multilabel classification
multi-task learning
multi-output regression
classification head
joint accuracy

Learn next

Write a small target table with two categorical columns and define each column's classes. Create example predictions and calculate per-output and joint correctness. Then compare modeling approaches while keeping the target definitions, data split and evaluation criteria fixed.

Sources

Scikit-learn developers, Multiclass and multioutput algorithms: https://scikit-learn.org/stable/modules/multiclass.html The four-request calculation, form analogy and support-team case are original educational illustrations. The documentation grounds the task distinction and independent-output strategy; no hypothetical result is presented as a measured model benchmark.

Previous: Multi-class logistic regression Next: Multi-document summarization