AI Glossary / AI and machine learning / Intermediate

Uncertain data

Definition

Uncertain data are data whose values, identities, relationships or existence are not known with complete confidence. Instead of treating every recorded item as an exact fact, an uncertainty-aware system represents and processes the relevant limits of knowledge.

What it means

A sensor may report a measurement with error, an extraction model may offer several possible readings and two records may or may not describe the same entity. These are different forms of uncertainty and may need different representations.

Uncertainty is not identical to missingness or poor quality. A carefully measured value with a documented uncertainty range can be more useful than an apparently precise value with unknown reliability. Missing data can also be one source of uncertainty, but filling a blank with one guess does not necessarily remove it.

In probabilistic data management, alternatives can be represented with probabilities and relationships between uncertain facts. Other settings use intervals, sets of possible values or explicit unknown states. The representation should match what is actually known; a probability should not be invented merely because a system expects a number.

How it works

The process starts by identifying what is uncertain and why. Measurement uncertainty, ambiguous identity and uncertain event occurrence have different implications. Metadata can record provenance, estimation methods and dependencies.

Queries or models then propagate the uncertainty according to a defined semantics. In a probabilistic database, a query may return a probability that an answer is true across possible data states. Dependence between facts matters: two records copied from one source are not independent confirmations.

Decision-making can use ranges, expected values or sensitivity analysis, depending on the task and consequences. Validation examines whether probabilities are calibrated or intervals have the intended coverage. Uncertainty should remain traceable through transformations so a downstream display does not turn a qualified estimate into an apparently certain fact.

Three practical examples

1. A location service represents a device position as an estimated region rather than one exact coordinate.

2. A document-extraction workflow preserves alternative readings and confidence information for an unclear field.

3. A record-linkage system marks a possible match as uncertain and routes consequential ambiguous cases for review.

Think of it this way

Think of a map with shaded regions where a location is plausible instead of a pin pretending to be exact. The limit is that uncertainty can concern whether an event happened or whether two records match, not only a numerical location range.

Worked example

Suppose two independent sensors each detect a specified event with probability 0.8 under a simplified model. The probability that at least one detects it is 1 - (1 - 0.8) squared = 0.96.

If the two reports are actually duplicates from the same sensor, treating them as independent would be wrong. Under perfect dependence, the probability of at least one detection remains 0.8, not 0.96.

The example shows why uncertainty is not only about attaching probabilities to individual rows. Relationships among those rows can change the result of combining them. The assumptions must be justified before the arithmetic is useful.

Illustrative case study

A fictional logistics team combines status messages from two systems and interprets agreement as strong confirmation. A review finds that both systems receive their messages from the same upstream scanner.

The team records the shared source and changes its confidence calculation. It also distinguishes a missing scan from a confirmed absence of the package, preventing the dashboard from treating incomplete evidence as a definite event.

The case illustrates the importance of provenance and dependence. More records do not necessarily mean more independent evidence, and a precise-looking confidence score can be misleading if its assumptions are hidden.

Common misconceptions and mistakes

• Uncertain data are unusable data. Explicit uncertainty can support better decisions than false precision.
• Two matching records always strengthen evidence independently. They may share a source or error mechanism.
• Replacing an unknown value with a single estimate removes uncertainty. It can merely hide it.

Business use

Uncertain-data methods are useful wherever measurements, extraction or integration produce imperfect information. They can help organizations avoid overconfident decisions and prioritize verification effort.

Identify the uncertainty relevant to the action and preserve its source. Use representations that downstream systems can interpret correctly, with clear distinctions between unknown, estimated and confirmed values. Evaluate calibration where probabilities are reported. Focus review on cases where uncertainty could change the decision rather than requiring identical treatment for every ambiguous field.

People also ask

Must uncertain data use probabilities?

No. Intervals, alternatives and explicit unknown states can be more appropriate when probabilities are unavailable.

Is uncertainty the same as randomness?

Not always. It can also reflect limited knowledge, ambiguous interpretation or incomplete measurement.

Why does provenance matter?

It helps reveal shared sources, update histories and dependencies that affect how evidence should be combined.

Check your understanding

Why should two identical reports copied from one source not be treated as two independent confirmations?

Answer: They share the same underlying evidence and potentially the same error, so multiplying or combining probabilities as if independent can overstate confidence.

Related terms

• Probabilistic database
• Calibration
• Missing data
• measurement uncertainty
• data provenance
• entity resolution
• uncertainty propagation

Learn next

Read data provenance and probability dependence. For a set of uncertain records, identify which are independent observations and which are copies or transformations of the same evidence.

More in AI and machine learning

• Tsetlin machine - A Tsetlin machine is a learning model that uses teams of Tsetlin automata to construct propositional-logic clauses.
• Tucker decomposition - Tucker decomposition represents a multiway array, or tensor, using a smaller core tensor and one factor matrix for each mode.
• Ugly duckling theorem - The ugly duckling theorem is a result associated with Satosi Watanabe showing that, if all possible predicates are treated equally in a particular logical...
• Underfitting - Underfitting occurs when a model fails to capture useful patterns in its training data and consequently performs poorly.
• Uniform convergence in probability - Uniform convergence in probability means that the largest error across an entire indexed collection of random quantities becomes small with high probability...
• Univariate feature selection - Univariate feature selection evaluates input features one at a time and selects those that score well under a chosen statistical criterion.

Sources

Nilesh Dalvi and Dan Suciu, Management of Probabilistic Data: Foundations and Challenges: https://homes.cs.washington.edu/~suciu/talk-pods2007-keynote.pdf . Independent review read uncertainty examples, probability and independence definitions, and possible-worlds query semantics. The sensor probabilities and fictional logistics case are original illustrations. Probabilistic databases illustrate one representation of uncertainty, not a requirement that every unknown receive a probability.

Previous: UNCANNY VALLEY Next: Uncertain database