TruthFinder

TruthFinder is a named approach from research on conflicting web information. It starts from the observation that trustworthy sources tend to provide credible claims, while claims supported by trustworthy sources in turn inform estimates of source trustworthiness. The method uses this mutual relation iteratively. It also considers relationships among values rather than treating every distinct string as wholly unrelated.

The approach illustrates an important practical distinction from raw voting: a claim's support depends on who reports it, while source trust is inferred from its broader claim history. Before applying that idea, claims need a common subject and property, and source identities must be meaningful. Copying, time changes and sparse coverage can still distort the apparent pattern of agreement.

For a defensible record, TruthFinder is a model whose assumptions and outputs require inspection, not an authority that can certify a value. The original claims and sources must remain available so a reviewer can understand a high-ranked result. Its relation to source reliability is reciprocal; its relation to majority voting is that it assigns unequal evidential weight to sources.

The first weakness is that agreement is rewarded whether or not it is independent. A group of sources that copy a single wrong value will give it a high score and then earn high trust from having reported it, so that copying is rewarded twice. The follow-up literature, including work on source dependence, builds on exactly this limitation. The second weakness is that a single trust score per source ignores the fact that reliability varies by property. The third is the starting point: because all sources begin equal, an early majority of copies can steer the iteration toward a self-consistent but wrong fixed point.

TruthFinder also assumes that one value is true for each property. For multivalued properties, such as the inventors of a patent or the countries in which a trial recruits, that assumption forces a choice where several answers are right. The survey by Li and colleagues compares TruthFinder with later methods on these dimensions and is the best single source for choosing among them.

TruthFinder is one point in a long line of methods. Majority voting is its simpler predecessor; the models that account for source dependence are its successors; and the Dawid–Skene model applies a similar alternating estimate to categorical labels from annotators. Large-scale systems that extract facts automatically add a further layer, separating the error of the extractor from the error of the source, as in Knowledge-Based Trust. For a reference catalogue the lesson is the shape of the argument and not the formula: trust and credibility are estimated together, and both estimates need independent checks.