Must the positive label be 1?

No. That is a common encoding, but labels can use other numbers or text. The mapping determines the meaning.

Can positive mean an undesirable event?

Yes. Faults, duplicate records, and spam are often the conditions being detected.

Can a multiclass model have a positive class?

A particular class can be treated as positive in a one-versus-rest evaluation. Specify which class and how results are aggregated.

AI Glossary / Causal inference and experimentation / Beginner

Positive class

Definition

The positive class is the category a classification task designates as the condition or event being detected. In a binary task, the other category is the negative class. If a model detects duplicate records, duplicate is the positive class and nonduplicate is the negative class. Positive names a role in the classification problem; it does not necessarily mean desirable, correct, favorable, or numerically greater than zero.

What it means

Classification reports often describe how well a system finds a particular condition. Calling that condition positive makes it possible to distinguish successful detections, missed detections, and false alarms consistently.

For a device-failure detector, failure may be positive even though it is undesirable. For a membership-renewal predictor, renewal may be positive and desirable. The choice comes from the question being modeled, not from the everyday emotional meaning of positive.

Specify the target carefully. “Cancels within 30 days” and “eventually cancels” define different events. A label can only be interpreted in relation to its scope, observation window, and evidence. These details matter more than whether the software stores the label as 1, true, or a word.

How it works

Define the positive condition and assign reference labels using an appropriate process. Train a classifier to distinguish that condition from its counterpart. Some models output a class directly; others output a score that a decision rule converts into a class.

For example, a workflow may predict positive when its score is at least a specified threshold. The score's meaning and threshold convention must be documented. Moving the threshold can change the number of positive predictions without changing the actual reference labels.

Evaluation distinguishes a true positive, which correctly identifies an actual positive, from a false positive, which predicts positive for an actual negative. A missed actual positive is a false negative. Precision asks how many positive predictions were correct, while recall asks how many actual positives were detected. Both depend on the declared positive class.

Three practical examples

1. Duplicate detection: a records system treats matching pairs as positive. A false positive can incorrectly suggest combining records that belong to different entities.

2. Maintenance screening: a model treats a defined equipment fault as positive so inspectors can distinguish detected faults from missed ones and unnecessary alerts.

3. Customer feedback: a classifier treats requests for a refund as positive, even when the message is polite or enthusiastic. The target concerns intent, not positive sentiment.

Think of it this way

Think of highlighting every document that contains a requested clause. Highlighted candidates are positive predictions because they claim to meet the search condition. Some may be incorrect, and some matching documents may remain unhighlighted. The analogy is limited because a model's decision may come from learned patterns and uncertain labels rather than an exact text search, so the highlight is not proof.

Worked example

Suppose a fictional evaluation contains 50 record pairs, of which 15 are genuine duplicates. Duplicate is the positive class. A model flags 12 pairs as duplicates, and 9 of those flags are correct.

There are 9 true positives and 3 false positives. Six genuine duplicates were not flagged, so there are 6 false negatives. The remaining 32 pairs are true negatives. Positive-class precision is 9/(9 + 3) = 75%, while recall is 9/(9 + 6) = 60%.

Now describe the same decisions with nonduplicate designated as positive. The new true-positive count is the old true-negative count, 32. The new false-positive count is the old false-negative count, 6. Precision for the newly designated positive class is therefore 32/38, about 84.2%. The predictions did not improve; the question and metric interpretation changed. Overall accuracy remains (9 + 32)/50 = 82% under either naming convention.

Illustrative case study

Fictional case: Firwood Support receives a vendor report stating “positive-class recall: 92%.” The internal team assumes positive means urgent tickets, but the vendor's configuration uses routine tickets as the positive class.

Before comparing systems, the team requests the label mapping and confusion counts, then recalculates the metrics for urgent tickets. It adds the target definition to every report and checks that probability columns correspond to the intended class. The case shows how a terminology mismatch can undermine interpretation even when the arithmetic is correct. It does not describe an actual vendor or verified performance result.

Common misconceptions and mistakes

Positive is not a synonym for true positive. A positive prediction becomes a true positive only when it agrees with the reference label. Nor does positive necessarily mean a majority class or an outcome that should be encouraged.

Do not assume that a model's first probability column belongs to the positive class. Software may order labels according to its own class mapping. Verify that mapping before applying thresholds or interpreting a score.

Another mistake is reporting precision or recall without naming the class of interest. In multiclass settings, each class may be treated as positive against the others for a particular calculation, and averaging methods summarize those results differently. State both the class definition and any averaging convention.

Business use

Clear positive-class definitions help operational teams understand alert volume, missed cases, and review workload. They also make vendor comparisons and model monitoring more consistent.

Include the event definition, time window, label mapping, and decision threshold in the model's evaluation record. When an event is rare, examine precision and recall alongside accuracy rather than relying on one headline number. If the business changes what counts as the target event, treat that as a meaningful evaluation change instead of a cosmetic rename.

Check your understanding

A model makes 30 positive predictions, and 24 match actual positive labels. How many false positives are present among those predictions?

Answer: Six. The other 24 are true positives. The number of false negatives cannot be determined without knowing how many actual positives were predicted negative.

Related terms

Learn next

Define a binary task and write its positive condition in one sentence, including any time window. Build a small confusion matrix, then reverse the positive-class designation and recalculate precision to see how its meaning changes.

Sources

Google, Machine Learning Glossary, positive class and related classification entries: https://developers.google.com/machine-learning/glossary#positive-class

The primary glossary grounds the class-role definition. The record-pair calculation, fictional reporting case, and exercise are original teaching material; all labels and counts are illustrative.

Previous: Positional index Next: Positive definite (pd)