IterativeImputer

IterativeImputer is a multivariate imputation estimator in scikit-learn that fills in missing values by modeling each feature with missing values as a function of the other features, in a round-robin fashion. It is inspired by the chained equations approach used in MICE (Multiple Imputation by Chained Equations) and is designed to exploit relationships among features to produce plausible imputations.

During fitting, the imputer iterates over features containing missing values. For each such feature, it trains a regression model using the other features as predictors on samples where that feature is observed, and then uses the trained model to predict the missing values. This process is repeated for a specified number of iterations to propagate information across features. The underlying regression model is provided via the estimator parameter and must implement fit and predict; the default is BayesianRidge, but other regressors (for example, random forests or gradient boosting) can be used.

Key parameters include max_iter, imputation_order, estimator, and sample_posterior. Imputation_order can be set to determine the order in which features are imputed (for example, ascending, descending, roman, or random). If sample_posterior is True, the imputer samples from the posterior distribution of the predictions, enabling multiple imputations; otherwise, it yields a single imputed estimate.

Usage considerations include that IterativeImputer is best suited for numeric data (categorical features should be encoded) and can be more computationally intensive than simple imputation methods. It is appropriate when there are complex relationships among features and missingness across multiple variables.