Imputation replaces missing values with estimated values or classes so tabular learning algorithms can use the data. It is especially useful when classical machine-learning methods cannot handle missing values directly. section 2.2.4
Choose the replacement strategy according to the available information and the meaning of missingness:
- Simple univariate imputation uses information from the same column, such as its mean or median. section 2.2.4 section 6.1.1
- Multivariate imputation uses the other available columns to reconstruct a more informed value or class. section 2.2.4 section 6.1.1
- A missing indicator can preserve the fact that a value was absent when missingness carries useful information. section 2.2.4 section 6.1.1
Comparing simple and multivariate estimates
A continuous-feature workflow creates 5% artificial missingness, fills one copy with SimpleImputer, and fills another with IterativeImputer using a RandomForestRegressor. The reported mean absolute errors are lower for multivariate imputation across the listed features, including number_of_reviews and availability_365.
