Glossary
Source dependence
Source dependence is the extent to which apparently separate sources derive claims from one another or a common upstream source.
Truth discovery
Source dependence occurs when two reports are not independent observations. One may copy another, both may syndicate a common feed, or both may inherit the same error from an upstream record. Agreement among dependent reports supplies less corroboration than agreement among genuinely separate observations. Dependence can vary by claim and period, so it should not be reduced to a single permanent relation between publishers.
In practice, analysts compare publication times, wording, cited origins and known distribution arrangements. A cluster of near-identical claims may indicate copying, but similarity alone does not establish its direction. Conversely, different wording does not establish independence if both reports draw from the same underlying document. Claims should retain their immediate source and, where known, their earlier origin.
Dependence matters whenever an organization must explain why repeated evidence was counted only once or given limited weight. Without that distinction, copied errors can look like consensus. The concept modifies source reliability and majority voting: a reliable publisher can still be repeating a mistaken upstream claim, and many votes can represent one original observation.
Example
Three directory pages repeat one filing about Halden Therapeutics AG (fictional); their agreement is treated as one lineage of evidence, not three independent filings.
Forms of dependence
Dependence takes several forms that call for different handling. Direct copying means one source took its values from another, sometimes with deliberate edits. Syndication means several outlets receive the same feed and publish it unchanged. Common upstream error means two sources each consulted the same primary record, which was itself wrong. Shared extraction means two datasets were produced by the same software run over the same documents, so that an extractor's systematic mistake appears in both. Only the first has a direction, and even there the direction is often unknown.
The practical test is not whether two sources agree but whether they could have disagreed. If one source cannot publish a value without the other having published it first, the later source adds no independent information about that value. It may still add information about other properties, which is why dependence is best recorded per claim and per period and not as a permanent relation between two publishers. Research on this problem, including the work on truth discovery with copying sources discussed on this site, models dependence as a probability that depends on how many false values two sources share, since sharing a rare wrong value is much stronger evidence of copying than sharing a common right one.
What to retain so dependence can be assessed
Dependence cannot be assessed from values alone. Retain, for each claim, the immediate source, the date it was observed, the origin the source itself cites if it cites one, and the exact wording. With those four items an analyst can group claims into lineages, order them in time, and see where one lineage fans out into many pages. The W3C PROV data model offers vocabulary for this: derivation between entities, attribution to agents, and the activities that produced each. A record that stores only the winning value and a source name has discarded what is needed to count evidence once.
A related defence is to store a retained copy of the source at the time it was read. Pages that appear independent when read today may turn out, from archived copies, to have been derived from a single document on a known date. Without the copy, that finding is impossible to reproduce.
A worked example
Four websites report that Halden Therapeutics AG (fictional) completed a licensing agreement on a given date. Three of them link to, or quote the opening sentence of, one trade-press article; the fourth is the counterparty's own investor page. Weighted by source count the claim has four supporters. Grouped by lineage it has two: the trade-press article with its three followers, and the counterparty's page. If the trade-press article itself rested on an unnamed informant, the first lineage deserves less weight than the second, and an analyst who notes that in the record has made the decision reviewable.
The test for the opposite error is equally important. Two sources that use similar wording because both quote an official filing are dependent on the filing but not on each other, and treating them as copies of one another would discard a genuine corroboration of how the filing was read. The article on fairness tests for truth-discovery aggregation and the one on sybil attacks show how deliberately manufactured copies can be used to move a naive aggregate, which is the adversarial form of the same problem.
Related terms and boundaries
Dependence is a refinement of two neighbouring ideas. It qualifies source reliability, since a reliable publisher that repeats an upstream mistake adds no new accuracy, and it qualifies majority voting, since a count of dependent sources overstates the evidence. It is usually established from provenance and data lineage when those exist, and inferred from patterns of shared errors when they do not. In both cases the finding is a probability and not a certainty, and the record should say which.
Checks a reviewer can run
Pick a value with many supporters and ask how many independent lineages stand behind it. Ask how copying is detected and what happens to a source once flagged. Check that each stored claim carries its immediate source, observation date and cited origin. Find a pair of sources flagged as dependent and read the evidence for the flag. Look for the opposite error too: sources treated as copies that in fact consulted the same primary document separately.
Common misreadings
- Counting syndication partners as separate witnesses.
- Inferring copying solely from matching values.
- Failing to retain the upstream source when it is known.
Appears in
- Majority voting vs truth discovery
A count of agreeing sources is a useful baseline; truth discovery also considers how those sources behave.
- Master data management vs knowledge fusion
MDM governs shared business entities; knowledge fusion reconciles assertions drawn from multiple sources.
- Estimating source reliability
Assess a source for a defined property and period using checked outcomes and visible uncertainty.
- Reporting source conflicts
Preserve competing values and their grounds until a scoped decision can be defended.
References
- Li et al. (2016), A Survey on Truth Discovery, ACM SIGKDD Explorations 17(2).
- Dong et al. (2014), Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion, KDD.
- W3C (2013), PROV-DM: The PROV Data Model.
- Dong et al. (2015), Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources, PVLDB 8(9).
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Source dependence" bear on a decision of yours?
This entry sits under truth discovery. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.