Is negative always encoded as zero?

No. Zero is common, but systems may use other values or text labels. Check the mapping.

Can there be several negative categories?

A binary task can combine several underlying categories into its negative class. Whether that grouping is useful depends on the question being asked.

Does a negative prediction mean zero probability?

No. It generally means the decision rule selected the negative class, potentially despite some estimated chance of the positive condition.

AI Glossary / Causal inference and experimentation / Beginner

Negative class

Definition

The negative class is one of the two classes in a binary classification task, defined as the counterpart of the positive class. If the positive class means that a message is spam, the negative class means that it is not spam under the task's labeling rules. Negative describes a classification role. It does not mean a negative number, an undesirable outcome, or missing information.

What it means

A binary classifier distinguishes two categories. Designers choose which category to call positive, usually because it represents the condition they want to detect. The other category becomes negative.

For a damaged-package detector, damaged might be positive and not damaged negative. For a successful-delivery predictor, success might be positive and failure negative. This reversal shows why the words positive and negative do not independently tell you whether an outcome is good or bad.

The label definition must be explicit. Unknown, unreadable, or unreviewed examples should not automatically become negative merely because nobody confirmed them as positive. That shortcut can introduce misleading training and evaluation data.

How it works

First define the two categories and a reliable process for assigning reference labels. A trained model may produce a class directly or a score related to the positive category. A decision rule then converts a score into one of the two predicted classes.

For example, a team might predict positive when a score is at least 0.70 and negative otherwise. The exact threshold and score meaning depend on the model and task. A below-threshold score does not prove that the event is absent.

Evaluation compares predicted classes with reference labels. A true negative is a reference-negative example correctly predicted negative. A false positive is a reference-negative example incorrectly predicted positive. A false negative instead belongs to the reference-positive class but receives a negative prediction. Keep actual class and predicted class separate when interpreting these names.

Three practical examples

1. Email filtering: when spam is positive, ordinary messages form the negative class. Incorrectly sending an ordinary message to spam is a false positive.

2. Quality inspection: when visible cracks are positive, inspected items without the defined crack condition are negative. Their labels depend on the inspection standard and available evidence.

3. Subscription analysis: if cancellation during the next month is positive, customers who do not cancel during that defined window are negative. A later cancellation does not retroactively change that earlier target definition.

Think of it this way

Think of sorting envelopes into a tray marked “requires postage correction” and another marked “does not require postage correction.” The second tray is the negative category because it lacks the selected condition. It is not the bad tray. The analogy is limited because a classifier may make uncertain predictions, whereas the tray labels alone say nothing about how reliably each envelope was assessed.

Worked example

Suppose a fictional detector reviews 200 parcels for a defined damage condition. Reference inspection identifies 40 damaged parcels and 160 undamaged parcels. Damage is the positive class, so the 160 undamaged parcels are the actual negatives.

The model flags 35 damaged parcels and misses 5. It also flags 12 undamaged parcels while correctly leaving 148 undamaged parcels unflagged. The confusion counts are therefore 35 true positives, 5 false negatives, 12 false positives, and 148 true negatives.

There are 153 negative predictions: 148 correct negatives plus 5 missed positives. This differs from the 160 actual negatives. Specificity, the fraction of actual negatives correctly recognized, is 148/160 = 92.5%. Negative predictive value, the fraction of negative predictions that are correct, is 148/153, about 96.7%. Different denominators answer different questions; neither number means that every unflagged parcel is undamaged.

Illustrative case study

Fictional case: Brookside Audio labels support requests as involving a known connection fault or not involving that fault. Its spreadsheet initially marks every unresolved request as negative. A reviewer notices that unresolved requests have not yet received a diagnosis and could belong to either class.

The team introduces an unresolved review status outside the training target and waits for sufficient evidence before assigning a binary reference label. It also documents the exact connection fault being detected, since “any product problem” would define a different positive class. The example illustrates label discipline, not an actual company's measured improvement.

Common misconceptions and mistakes

Do not confuse the negative class with negative sentiment. In a sentiment classifier, a designer could assign the positive detection role to critical comments, making negative sentiment the positive class of that particular binary task.

The negative class is not necessarily the majority class. A carefully sampled dataset may contain equal numbers, and some settings contain more positives. A numerical code such as 0 is also only an encoding convention; inspect the class mapping instead of assuming it.

Finally, predicted negative is not equivalent to confirmed absence. Prediction errors, ambiguous labels, and changing conditions can all affect the conclusion. Report the model's decision and the available evidence at the appropriate level of certainty.

Business use

Understanding the negative class helps businesses interpret screening, routing, and alert systems. Teams need to know how many ordinary cases are unnecessarily escalated as well as how many important cases are missed.

Write the positive and negative definitions alongside reports, including the observation window and exclusions. When evaluating a supplier, ask how unreviewed cases were handled and whether class proportions resemble the proposed operation. A seemingly strong result can reflect a dataset dominated by easy negatives rather than useful detection of the target event.

Check your understanding

A model predicts 80 examples as negative, and 6 of those examples are actually positive. How many of the 80 are true negatives?

Answer: There are 74 true negatives and 6 false negatives. The total number of actual negatives cannot be determined without knowing the remaining predictions and labels.

Related terms

Learn next

Choose a familiar two-category task and write exact class definitions. Make a small confusion matrix with 20 examples, then count actual negatives separately from negative predictions. Explain why the totals can differ.

Confusion matrix

Sources

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

The primary glossary supports the binary-class distinction. The parcel arithmetic, labeling case, and exercises are original teaching examples. Class encodings, thresholds, and handling of unknown labels must be checked for each implementation.

Previous: Negation Next: Negative-free representation learning