AI Glossary / AI and machine learning / Intermediate

Upweighting

Definition

Upweighting increases the contribution of selected examples to a calculation or learning objective. In the supplied glossary's sampling context, it compensates for downsampling by weighting retained examples according to the factor by which their class was reduced.

What it means

Large datasets sometimes contain far more examples of one class than another. A team may retain only a fraction of the common class to reduce processing. The resulting sample no longer has the same class proportions as the original collection.

Upweighting can account for that deliberate sampling. A retained example stands for more than one original example in a weighted calculation. The appropriate factor depends on its inclusion probability.

The term is also used more broadly when increasing the influence of important or difficult cases. That broader use changes the learning emphasis and should not be confused with correcting a known sampling design.

For business readers, weights explain why 100 retained records may count as more than 100 units of evidence in a calculation. They do not create additional independent observations or recover every detail discarded by sampling.

How it works

The team records how examples were selected. If each example in a class had an equal probability p of being retained, inverse-probability weighting uses a factor of 1 / p for that class's retained examples.

The weight is applied where the method supports it, such as in a loss function or an aggregate estimate. Some implementations normalize weights, so the exact interpretation of numerical totals must be checked.

If sampling probabilities differ across examples, one class-wide factor may be inadequate. Unknown or selective inclusion cannot be corrected reliably by choosing a convenient multiplier.

Evaluation should reflect the intended population and decision. Weighted training does not automatically produce calibrated probabilities or eliminate all effects of a changed sample, so those properties need direct assessment.

Three practical examples

• Event classification: retain a fraction of abundant routine events and account for that selection in the training objective.
• Survey-style analysis: use known inclusion probabilities when estimating totals from a deliberately sampled collection.
• Model comparison: calculate weighted metrics when a review sample oversamples rare cases but the desired summary concerns the original population.

Think of it this way

Think of sampling every tenth box from a uniform shipment and letting each inspected box represent ten boxes in a summary. The multiplier reflects the sampling plan.

The analogy has a limit. If workers preferentially choose boxes that look undamaged, multiplying by ten does not fix that selection bias. Weighting works only as well as the assumptions and recorded inclusion probabilities behind it.

Worked example

Restoring a sampled class's contribution

Suppose an invented dataset contains 9,000 routine events and 1,000 exceptional events. The team retains one tenth of routine events, leaving 900, and keeps all exceptional events.

Assigning weight ten to each retained routine event gives a weighted routine count of 900 times 10 = 9,000. Exceptional events keep weight one, giving 1,000. The weighted proportions match the original 90% and 10% class proportions.

The actual training file still contains only 1,900 records. Weighting has not restored the 8,100 discarded routine examples or any unusual patterns they contained.

The calculation assumes the routine sampling was uniform with the stated retention rate. It illustrates a sampling correction, not a guarantee that a particular classifier will behave identically to one trained on every original record.

Illustrative case study

A fictional monitoring team samples routine events more heavily from quiet hours than busy hours but applies one constant weight to all retained routine records. Its summary underrepresents busy-period behavior.

The analysts document the selection process by time period and use appropriate inclusion information where available. They compare results with a fresh representative sample and reconsider whether the sampling design serves the task.

The lesson is that a weight must correspond to the actual selection mechanism. A simple multiplier cannot repair an undocumented or misunderstood sampling process.

Common misconceptions and mistakes

• Upweighting creates more independent data. It changes influence, not the number of observed cases.
• Any class imbalance should be removed. The desired objective and real population determine whether and how weighting is appropriate.
• One multiplier fixes every sampling bias. Different or unknown inclusion probabilities require more careful treatment.

Business use

Upweighting helps teams manage large datasets and interpret deliberately balanced review samples. It makes sampling decisions visible in the learning or measurement process.

Ask why weights exist, how they were calculated and which operations actually use them. Report both raw sample counts and weighted summaries. Check performance and probability calibration on data representing the intended use, rather than assuming the weighting scheme solves every downstream issue.

People also ask

Is upweighting the same as duplicating records?

It can have a similar effect in some objectives, but implementation and optimization details can make them differ.

Must weights be whole numbers?

No. Inverse-probability and other weighting schemes can use non-integer values.

Does weighting always improve accuracy?

No. It changes the objective or estimate; whether that helps depends on the task, data and evaluation metric.

Check your understanding

A class is uniformly retained at one fifth of its original size. What inverse-retention weight represents its original contribution?

Answer: Five. Each retained example represents five under that sampling design, but the weight does not recreate the discarded records.

Related terms

• Calibration
• downsampling
• sample weight
• class imbalance
• importance weighting
• weighted loss

Learn next

Read downsampling and sample weight. For one sampled dataset, document inclusion probabilities and compare raw counts with weighted counts before interpreting any reported metric.

More in AI and machine learning

• Unsupervised Clustering - Unsupervised clustering groups examples by patterns in their inputs without using predefined target labels to determine the groups.
• Unsupervised learning - Unsupervised learning finds structure in data without using supplied target labels to guide the fitting objective.
• Unsupported-claim rate - Unsupported-claim rate is the proportion of evaluated claims in a response that lack adequate grounding under a defined evidence standard.
• VACUUM - VACUUM is a data-quality mnemonic referring to a Valid, Accurate, Consistent, Uniform and Unified Model.
• Value imputation - Value imputation replaces missing data values with chosen or estimated substitutes.
• Variable kernel density estimation - Variable kernel density estimation estimates a probability density using smoothing kernels whose bandwidths vary across observations or locations.

Sources

Google, Machine Learning Glossary, upweighting: https://developers.google.com/machine-learning/glossary#upweighting. The downsampling-compensation sense was consulted. Sampling arithmetic and monitoring scenario are original.

Previous: Upper ontology Next: URL normalization