Multivariate imputation replaces missing values by using relationships among multiple available features rather than relying only on one column. It is especially useful when missingness is completely at random (MCAR) or depends on observed variables (MAR), because correlated predictors can help reconstruct plausible values. section 2.2.4 section 6.1.1
A practical workflow compares simple and multivariate approaches before choosing a preprocessing strategy:
- Identify whether missingness may itself carry information; add a missingness indicator when the absence of a value reflects how the data was generated. section 2.2.4
- Use simple imputation as a baseline, such as replacing values with an expected value, mean, or median. section 6.1.1
Comparing replacement quality
For continuous features such as minimum_nights, number_of_reviews, reviews_per_month, calculated_host_listings_count, and availability_365, the multivariate approach produces lower reported mean absolute errors than simple imputation for every listed feature.
Models including XGBoost, LightGBM, and HistGradientBoosting can handle missing values natively, but direct imputation may still make feature handling easier to inspect. Missing-not-at-random cases require specialized treatment and are not addressed here. section 2.2.4 section 6.1.1
