Conflicting values across sources and how to pick one without guessing
Conflicting values across sources are settled by measuring, per field, how often each source agrees with a confirmed value. Rank the sources separately for every attribute, because a lead capture app can be the best source of phone numbers and the worst source of job titles in the same file.
One buyer, three job titles. The web registration says VP Marketing. The lead capture scan from an exhibitor stand says Marketing Director. The exhibitor portal, where a stand contact typed her in as a guest, says Head of Marketing.
Nobody is lying and nothing is broken. Conflicting values across sources are the ordinary output of collecting the same fact through four routes over five years, and the question is only which one goes on the record and on what basis.
The usual basis is whichever system somebody decided was the master, and that decision is almost always made once, globally, for every field at the same time. That is the part worth reconsidering.
What kind of conflict is this?
Rahm and Do catalogued the varieties in the IEEE Data Engineering Bulletin in 2000, and their classification is still the most useful map of the ground. They separate single source from multi source problems and schema level from instance level ones, and their example of a contradicting record is exactly the case above: two rows where "the same real world entity is described by different values".
They also list the multi source instance problems that make an event stack messy in ways that look like conflicts and are not. Different value representations for the same attribute. Different interpretations of a value, their example being currency units. Data provided at different aggregation levels. And records that refer to different points in time.
Sort your conflicts into those buckets before writing any rule, because two of them have no arbitration in them at all.
Normalise before you arbitrate
Look again at VP Marketing, Marketing Director and Head of Marketing. Depending on the company, those may be one role written in three title conventions, an American form and two British ones. If so, there is nothing to arbitrate. The correct output is one normalised title, and picking a winner between the surface strings is answering a question that should not have been asked.
Run the values through title normalisation first and count how many conflicts survive it. Whatever is left after that step is the real arbitration problem, and on most files it is a good deal smaller than the raw disagreement count implies. The normalisation itself is job title normalisation in K9.
The same applies to company name, country, and any field with a coding standard behind it. Arbitration is for the residue, and the residue is smaller than the raw conflict count suggests.
Some of it is a schema problem wearing a data costume
Rahm and Do separate naming conflicts, where the same name is used for different objects or different names for the same object, from structural conflicts, where the same object is represented differently across systems. Both turn up in an event stack disguised as disagreeing values.
The clearest case is a company field. In a registration table it holds the visitor's employer. In a lead capture export it sometimes holds the exhibitor doing the scanning, because the export was designed for the exhibitor's own CRM. Load both into one column and you have a conflict that no arbitration rule can settle correctly, since one of the two values is an answer to a different question.
Country behaves the same way. Country of residence, country of the billing address, country of the event and country inferred from a phone prefix are four different attributes that end up in one column called country more often than anyone admits.
The test is cheap. For any field where two sources disagree a lot, pull 20 disagreeing rows and read them. If the pattern is that one source is consistently describing something else, you have found a mapping error, and the fix is in the ingestion layer rather than in survivorship. Fixing it there removes the conflict permanently, while ranking the sources merely picks a winner between two answers to different questions, forever.
A useful signal for this: genuine data conflicts are scattered, and mapping errors are systematic. A source that disagrees on 4 per cent of rows is noisy, and one that disagrees on 70 per cent is answering something else.
Rank each source separately for every field
Here is the measurement that changes how the rest of the work is organised.
Take a reference set of contacts where the value is confirmed, which for most organisers means a pre show confirmation step, a badge collection check or a delegate correcting their own record. Suppose you have 480 such contacts. For each field, count how often each source agreed with the confirmed value.
On job title, the web registration agreed on 413 of the 480, which is 86 per cent. The lead capture app agreed on 293, or 61 per cent. The exhibitor portal agreed on 355, or 74 per cent.
On phone number, the ordering inverts. The lead capture app agreed on 437 of 480, or 91 per cent, because a rep scanning a badge and asking for a mobile gets a working number. The web registration managed 355, or 74 per cent, and the exhibitor portal 302, or 63 per cent.
On company name, the exhibitor portal wins with 466 of 480, or 97 per cent, since the stand contact knows the legal name of the company they are inviting guests on behalf of.
Three fields, three different winners. A single global ranking would have picked one system and imported its weakest field into the golden record, and the failure would be invisible because the record would look complete.
So the ranking takes the shape of a small matrix: source down one side, field across the other, with a measured agreement rate in each cell. It fits on one screen for most portfolios, it takes an afternoon to compute once the reference set exists, and it is the artefact that makes every subsequent survivorship argument short. The rules that consume it are survivorship rules in J26.
How much of a difference is a real difference?
Agreement rates from 480 records carry sampling error, and the matrix invites over reading.
The standard error on a proportion near 0.86 from 480 observations is the square root of 0.86 times 0.14 divided by 480, which is about 0.016. Two standard errors is roughly three percentage points, so 86 per cent and 84 per cent are the same number, while 86 per cent and 61 per cent are decisively different. Rank the cells into bands rather than into positions, and only prefer one source over another when the gap is several times the interval.
Where two sources sit in the same band for a field, do not use the matrix to separate them. Fall through to another rule, usually recency, which is where the most recent wins rule in J28 belongs in the ordering.
Recompute the matrix annually and after any platform change. A registration form that adds a job title dropdown will move that source up a band within one cycle, and a rule matrix that reflects 2023 collection practice is quietly wrong about 2026.
What to do when you have no reference set
Most organisers do not start with 480 confirmed records, and the temptation is to fall back on the majority. Three sources say Marketing Director and one says VP Marketing, so Marketing Director wins.
Treat that with suspicion. Dong, Berti-Equille and Srivastava, in the Proceedings of the VLDB Endowment in 2009, put it plainly for exactly this situation: "Straightforward ways to resolve such inconsistencies (e.g., voting) may lead to noisy results, often with detrimental consequences." Their concern was sources that copy from one another, and an event technology stack copies constantly. A lead capture app that pre fills from the registration record and an email platform synced from the same registration record are one vote wearing three badges.
Two practical moves follow. Map the copying before you count votes, so that any set of systems fed by the same upstream table votes once. And build the reference set incrementally: every time a delegate corrects their own record, that correction is a confirmed value, so capture it with the field name, the old value, the source that supplied it and the date. A year of that produces the matrix for free, and it produces it from your own audience rather than from an assumption.
Where this stops
The uncomfortable case is when all three values were true.
VP Marketing in 2023, Head of Marketing after a restructure in 2025, Marketing Director at a different employer in 2026. No arbitration rule recovers that, because the conflict is an artefact of flattening a history into a single field. Where a field has a real timeline, the honest model is a set of dated observations with the current value derived from it, and the cost is that every query that reads the field becomes more complicated. Choose deliberately, and know that keeping the field scalar is itself a decision to discard the history.
The reference set has its own bias. People who confirm their details at a pre show check are more engaged than people who do not, and their records are cleaner to begin with, so agreement rates measured on them will overstate every source. The ranking survives that bias reasonably well, since it is comparative and the bias applies to all sources at once, and the absolute rates should be treated as optimistic.
One more limit worth naming. A source can be accurate and useless if it is accurate about something else. An exhibitor portal that records the company the guest was invited by, rather than the company they work for, will score well against a confirmed company name for most contacts and be systematically wrong for consultants and agency staff, which is resolving contacts to companies in J25.
Pick one field this week, ideally job title, and pull every value your systems hold for 100 contacts who attended more than one show. Count how many of the 100 have more than one distinct value after normalisation. That number is the size of the arbitration problem you actually have, and it is the first thing worth knowing before any unified data rule gets written.
Questions people ask about conflicting values across sources
- How do you resolve conflicting values between two data sources?
- Measure before you decide. Build a small reference set of records where the person confirmed their own details, then count how often each source agreed with the confirmed value for each field. Use those rates as the ranking, and apply the ranking per field rather than choosing one system as the master for everything.
- Is majority voting a good way to pick between conflicting values?
- It is weaker than it looks. Dong, Berti-Equille and Srivastava noted in 2009 that straightforward approaches such as voting can produce noisy results when sources copy from one another. In an event stack, a lead capture app that pre fills from the registration record is not casting an independent vote.
- Should one system be the master source for all fields?
- Rarely. Measured agreement rates usually cross over between fields, so the system that is best on company name is often not the best on phone number. A single global ranking then imports the weaknesses of the winning source into every field where it happens to be poor.
Related reading
- The most recent wins rule and the two fields it gets wrong
- Survivorship rules that decide which value wins on a merged attendee record
- Resolving contacts to companies when the badge says a domain you never saw