Dataset Value Is a Relationship, Not a Number

“How much is this dataset worth?” gets no stable answer, because a dataset has no intrinsic value. It has value relative to a specific use: the same file can be the difference between a working model and a stalled project for one team, and a maintenance liability for another.

Treat value as a function of two things: the data and the use, Q(D, U). Quality is not a property of the data alone. It is a relationship between the data and a job, and that framing changes how you evaluate a dataset before you pay for it or build it.

Quality is fitness for use, not a score

The classic definition comes from Joseph Juran: quality is fitness for use. For data, that means a dataset is high quality when it meets the needs of the people using it for a specific task. The same dataset can rate excellent for one team and fail for another because the use changed, not the data.

The six dimensions to check

The ISO 8000 series, the international standard for data quality and master data, and DAMA-DMBOK, the data management body of knowledge, use different words for the same checklist. In practice, teams settle on six dimensions:

  1. Accuracy. The data matches the real-world fact it claims to record.
  2. Completeness. Nothing expected is missing.
  3. Consistency. The same fact agrees across records and systems.
  4. Timeliness. The data is current enough to act on.
  5. Validity. The data conforms to the expected formats and rules.
  6. Uniqueness. There are no duplicates where there should not be.

None of these is a pass/fail number that stands alone. The acceptable bar for each one depends on the use. A field that is missing a third of its middle names is fine for a marketing list and fatal for a join key. Completeness only means something relative to the job.

Value depends on use even when quality passes

A dataset can clear all six dimensions and still be the wrong asset, for reasons that have nothing to do with the rows:

  • Acquisition cost. What it costs to get, set against what it would cost to build or buy elsewhere.
  • Recoverability. What it would take to get the data back after corruption or loss.
  • Provenance. Where it came from, and whether you can legally and ethically use it the way you intend.
  • Bias. Whether the population it samples matches the population you will apply it to.

An accurate dataset drawn from the wrong population is a liability that looks like an asset.

When data feeds a model, value is marginal contribution

Once a dataset is training material, its value shows up as marginal contribution to model performance. Data Shapley, introduced by Amirata Ghorbani and James Zou in 2019, turns that into a number: it values each data point or source by how much it improves performance on a fixed test set, averaged fairly across all combinations of data. The same data can be highly valuable to one model and nearly worthless to another, because the model and the task decide that value, not the data alone.

You cannot rank datasets on a spreadsheet and take the top one. The evaluation has to happen against the use, the model, and the decision the data feeds.

For AI work

Most data work starts as a harmonization problem: multiple sources, different schemas, uneven quality, all fused toward one use. Someone has to know what the use actually requires before “good data” means anything. That knowledge is domain expertise, and no generic quality score replaces it.

Before you buy or build a dataset, write the use down. Run the six dimensions against it, check the cost, recoverability, provenance, and bias, and where a model is involved, measure marginal contribution. The number you end up with is only meaningful against that use.