Learn AIExamples are measurements.
Inspect missing values, copied requests, unit mismatches and conflicting labels before treating a table as evidence.
By the end: Distinguish features, labels, observations and sampling units; explain how measurement choices alter a summary.
Start with something concrete.
A spreadsheet has six rows. Does that mean six independent requests, six different people or six reliable labels? Before calculating a score, inspect what each row actually records and how it entered the table.
Try the interactive model ↓A blank “hours until event” cell is replaced with zero. What assumption has been added?
Change one thing. Watch what follows.
- Inspect the original rows. Find the copied request, missing duration, mixed units and two label disagreements.
- Convert minutes to hours, then remove the copied row. Trace how each separate action changes the sum and denominator.
- Compare omitting the unknown duration with inventing a zero. Explain which population each average can reasonably describe.
- Count unique request IDs and compare annotator totals with row-by-row agreement. Propose a definition or clarification that could resolve disagreement.
Loading this browser experiment…
Rows come from a collection process
A feature is input information such as time until an event. A label records a target such as whether the request is urgent. An observation is a recorded example. The sampling unit is the thing you intend to count or make claims about: a request, person, site or time period. These are not interchangeable.
The lab’s six fictional rows contain five request IDs because request A was copied into another export. Five requests still need not mean five people. One person could submit several requests. The sample is the observed collection; the population is the wider set you hope to describe. Neither the number of rows nor their neat appearance proves representative coverage.
Units change the meaning of a number
Row 2 records 180 minutes while other rows use hours. Adding these raw values produces a number with mixed units, not a meaningful average duration. Convert before comparing:
The conversion factor is one hour per sixty minutes. The minute units cancel, leaving hours. After conversion and removal of the known copied row, the observed durations are 2, 3, 48 and 24 hours. The missing duration remains unknown.
The numerator is a sum in hours and the denominator is the count of observed entries. This is the mean of the four observed durations, not automatically the mean for all five requests or for future requests.
A blank is not zero
Zero hours means the event is now. A missing deadline does not say that. If you insert zero, the arithmetic becomes 77 / 5 = 15.4 hours because you have invented a fifth observation. The workbench makes this consequence visible instead of treating filling as a harmless formatting choice.
Omitting a missing value is also a decision. If people with urgent or confusing requests are more likely to omit the deadline, the observed mean can be unrepresentative. Investigate how values go missing, document the policy and avoid unsupported conclusions. Later, when fitting preprocessing for a model, use training data only to estimate any filling values.
Repeated rows change weighting
Keeping the copied A row counts the two-hour request twice. After unit conversion, the mean becomes (2 + 3 + 48 + 2 + 24) / 5 = 15.8 hours. This difference is caused by extra weight on one request, not by new evidence about the population.
The lab identifies one exact copied export. In real data, similar feature values do not prove two records are duplicates. Different requests may happen to look identical, and repeated measurements may be intentional. Inspect identities and provenance before deleting or merging records. Group related observations appropriately when evaluating a model.
Label agreement is about the same examples
Annotators A and B label the same total number of requests urgent, yet disagree on rows 2 and 6. A total discards which requests received the label. Compare row-level labels and ask whether the target definition, available context or measurement process explains the difference.
More voting may make a procedure more consistent but does not automatically reveal a single objective truth. Document uncertain or disputed cases, preserve the original evidence and seek an appropriate resolution process. Repairing units cannot resolve disagreement about what “urgent” should mean.
The workbench is a measurement audit, not model training. Use it to distinguish units, missingness, repeated observations, label definitions and sampling coverage before moving to the data-quality and leakage lesson.
Two annotators each label 2 of 5 requests urgent. Does that prove they agree on the labels?
Take this idea with you.
A row is an observation created by a process. Name the sampling unit, preserve unknowns, reconcile units and inspect disagreement before fitting a model. Splitting and train-only preprocessing come next.