Missing values are ordinary: a customer skipped an optional field, a sensor dropped out for an hour, a system migration left a column empty for older records. Many models cannot accept empty cells, so a team has to drop the rows, drop the column, fill the gaps or choose a model that accepts missing values.
Filling a gap can erase a signal. The scikit-learn guide to imputation notes that preserving the information about which values had been missing can be informative, and its imputers can add an indicator column for that purpose. A missing income field, for instance, may say something about an applicant that a filled-in average does not.
Common methods:
- a constant or a column statistic, such as the mean, median or most frequent value
- a value predicted from the other fields in the same record, known as multivariate imputation
- a value taken from the most similar records, as in nearest-neighbor imputation
- an indicator column that records which values were missing, kept alongside the filled-in value
How is data imputation different from deleting missing data?
- Deleting missing data removes incomplete rows or columns. The data keeps only observed values but can shrink, and can lean toward the records that happen to be complete.
- Data imputation keeps the rows and replaces gaps with estimates. The sample size is preserved, but the data now contains values nobody observed.
- Which choice is safer depends on why the values are missing. Statistical treatment of that question builds on Donald Rubin’s 1976 Biometrika paper, Inference and Missing Data.
Where does data imputation break in AI work?
Data imputation breaks when filled-in values are treated as observed. A column that was mostly empty and then filled with its median looks complete in a data profile, and a check that counts missing values will score it as clean. The score improves, while the model can lose the information that those values were missing, which may itself be a signal.
Imputation can also leak information when the imputer is fitted on all the data before the train-test split, because the filled-in values then carry information from the test records, a form of data leakage. The write-up of CUBIG’s UCI Shill Bidding experiment (October 2026), for example, states that its script fits the imputer, along with the scaler and classifier, using training data.
Fitting the imputer on training data only, keeping a missing-value indicator, and recording which method was used on which data version, and what it preserved, lets a team compare alternatives, such as dropping the column or leaving the gap marked, under the same evaluation conditions.
Not every gap should be imputed. In sales records, values that people outside the model read as facts, such as order amounts and customer IDs, are better flagged or left blank; missing data in sales records works through that decision field by field, with figures from the UCI Online Retail dataset.