Profiles inferred from raw data are often absent from dashboards
Aliases: inferred-data visibility · profile transparency · inferred attribute disclosure
What it is
Inferred profile visibility extends dashboards beyond supplied or directly observed records to classifications, scores, interests, risk labels, and predictions computed from them. A profile is not more raw data; it is a fallible system claim about a person that may influence ranking, pricing, review, or advertising more directly than its inputs.
Why it happens
Data engineering often defines user data as exportable source tables while treating features, segments, and intermediate scores as internal state. People then see records that produced a judgment but not the representation actually used for differential treatment. Profile visibility exposes misclassification, stale inference, and unexpected secondary use and gives challenge a concrete object.
Studying it
Create accounts with known events, features, labels, and downstream decisions. Ask participants to find what the system thinks, judge evidence and effect, and correct one planted error. Reconcile the dashboard by lineage with feature stores, model output, and decision logs, measuring coverage, uncertainty understanding, correction success, and downstream result. Showing only interest labels does not establish coverage of high-impact scores.
Where it stops holding
Exact fraud and abuse features or thresholds may enable evasion; category, main basis, impact, and appeal can remain visible. Outputs may be transient or probabilistic and should not be stated as stable personality facts. Group-level statistics also differ from individualized judgments and should not imply certainty about one person.
Applying it
- Inventory features, segments, scores, labels, and predictions, prioritizing those affecting function, display, price, risk, or human decisions.
- Show a readable name, current value, uncertainty, principal sources, update time, and actual effect.
- Offer challenge, correction, restriction, or appeal and state whether action changes source data, profile, or downstream decision.
- Generate coverage from feature stores and decision logs; do not claim completeness while a hidden output still affects the person.