AI Glossary / Data engineering and architecture / Intermediate

Truth discovery

Definition

Truth discovery is a family of data-integration methods that estimates which claims are most likely to be correct when multiple sources provide conflicting information. Many methods jointly estimate source reliability and the credibility of candidate values. The name describes inference, not guaranteed truth from unreliable or incomplete evidence.

What it means

Suppose several catalogs disagree about a product's weight. Simply taking the most common value treats every catalog as equally informative. A truth-discovery method may instead consider that some sources have a stronger record of accurate information than others.

The challenge is that reliability itself may be unknown. Many methods infer it from patterns across multiple objects and claims. This creates a relationship between the selected values and estimated source quality. The result depends on assumptions, available evidence and the method used; a high internal score should not automatically be interpreted as a calibrated probability of correctness.

How it works

The system aligns claims referring to the same object and property, normalizes representations such as units and records their sources. Without this preparation, two compatible statements may appear contradictory or claims about different objects may be compared incorrectly.

A common iterative approach starts with provisional source weights, estimates likely values and then revises source weights according to their agreement with those estimates. It repeats until its stopping condition is met. Other approaches use probabilistic or optimization-based formulations.

The model must fit the task. Some properties have one value at a given time; others legitimately have several. Sources may copy one another, values may change and reliability may differ by topic. Convergence only means the procedure reached its stopping condition. It does not independently establish that the selected claim is true.

Three practical examples

1. Product information: A catalog aggregator reconciles conflicting dimensions while preserving original sources, units and product-version identifiers. Unresolved disagreements can be routed for review.

2. Sensor integration: Several sensors estimate the same physical quantity. A method considers their estimated reliability rather than assuming every reading deserves equal influence, while checking whether all sensors share the same calibration problem.

3. Knowledge-base maintenance: A team compares claims from multiple documents about an organization's attributes. It records dates and provenance so an older correct statement is not automatically treated as an error about the present.

Think of it this way

Think of an editor comparing reports from several correspondents. The editor considers their evidence and previous performance, while checking whether several reports merely repeat one original account.

The analogy has limits. An automated method usually works with a formal model and limited signals, not the full judgment or investigation available to an editor. Agreement can arise from shared mistakes, and a source reliable on one topic may be weak on another. The inferred result remains open to new evidence.

Worked example

Suppose three sources report a component's mass. Source A says 10 grams, while B and C say 12 grams. A simple majority selects 12. For illustration, assume source weights A = 0.8, B = 0.3 and C = 0.2, estimated from other relevant evidence.

The support for 10 is 0.8, while the combined support for 12 is 0.5. This illustrative rule selects 10 despite its minority support. These weights are demonstration scores, not independently established probabilities, and this calculation is not a complete truth-discovery algorithm.

Now discover that A describes an older component revision, while B and C describe the current revision. The original conflict was incorrectly framed: the claims concern different versions. Better weighting cannot repair that matching error. Correct entity, property and time alignment must come before the aggregation result is interpreted.

Illustrative case study

Fictional case study: EvidenceMeadow gathers specification claims from supplier pages and reseller catalogs. Its initial process treats each page as an independent source. Several reseller pages repeat the same incorrect measurement, giving that value strong apparent support.

The team records where claims originated and investigates likely copying relationships. It also keeps the original unit, revision date and supporting passage with each claim. Cases with insufficient independent evidence remain flagged instead of being converted into confident catalog values.

EvidenceMeadow evaluates changes against a checked sample that was not used to tune the method. It learns that counting sources, estimating reliability and verifying facts are different activities. A more sophisticated aggregation model still needs appropriate evidence and a way to handle uncertainty.

Common misconceptions and mistakes

A common mistake is equating popularity with correctness. A second is assuming that every source is independent merely because it has a different address. Repeated copying can make one assertion look like many confirmations.

Do not force a single answer when the property allows several valid values or changes over time. Avoid treating source reliability as a universal score. A technical catalog may be reliable for dimensions and poor for current availability. Finally, an algorithm's confidence or agreement score needs empirical evaluation before it is used as a probability or decision threshold.

Business use

Truth discovery can support data integration, catalog maintenance and knowledge-base construction when conflicting claims would otherwise require extensive manual reconciliation. It can prioritize review and provide a consistent way to combine evidence.

Businesses should define the relevant objects, properties and time scope, preserve provenance and evaluate against independently checked examples. Measure errors across source types and difficult cases rather than only overall agreement. For AI systems using the resulting data, retain uncertainty and alternative claims where appropriate. A generated answer should not become more authoritative merely because its input passed through an aggregation algorithm.

People also ask

Is truth discovery the same as majority voting?

No. Many methods estimate source reliability and use it when combining claims, rather than giving every source equal influence.

Does agreement among many sources guarantee correctness?

No. Sources can copy one another or share an underlying error.

Can the correct answer change over time?

Yes for time-dependent properties. The task must distinguish a changed value from a mistaken claim about the same time.

Check your understanding

Ten websites repeat one supplier's claim. Should their agreement automatically count as ten independent confirmations?

Answer: No. Their dependence matters. The repeated claim may have only one underlying origin, so the evidence should not be interpreted as ten independent observations.

Related terms

• Data fusion
• source reliability
• provenance
• entity resolution
• knowledge graph
• conflicting data
• confidence calibration
• fact-checking

Learn next

Collect conflicting claims about one harmless technical property. Normalize units, distinguish product versions and trace original sources before trying any voting or weighting rule.

More in Data engineering and architecture

• Tree contraction - Tree contraction is an algorithmic technique that repeatedly reduces a tree to a smaller structure while preserving information needed to solve a problem.
• Trivial Functional Dependency - A trivial functional dependency is a dependency X → Y in which every attribute of Y is already included in X.
• Trivial MVD - A trivial multivalued dependency, or trivial MVD, is a multivalued dependency X →→ Y that holds automatically because Y is contained in X, or because X...
• Tuple Constraint - A tuple constraint is a data integrity rule that can be evaluated on each tuple, or row, independently of other tuples.
• Tuple-oriented approach - A tuple-oriented approach handles database records one tuple or row at a time at the application or interface level.
• Tuple space - A tuple space is a shared coordination space in which concurrent processes exchange structured collections of values called tuples.

Sources

Yaliang Li and colleagues, “A Survey on Truth Discovery”: https://www.kdd.org/exploration_files/Article1_17_2.pdf

“Truth Discovery Algorithms: An Experimental Evaluation”: https://arxiv.org/abs/1409.6428

“On the Discovery of Evolving Truth,” indexed research text: https://pmc.ncbi.nlm.nih.gov/articles/PMC4688022/

The survey and evaluation abstract were opened. The temporal study was available in indexed text; direct access displayed a browser check. Weights and EvidenceMeadow are fictional illustrations, not validated predictions.

Previous: Truth Next: Truth-revealing