Benchmarks

This page summarizes benchmark results for the MultivariateImputer and the TimeSeriesImputer across multiple open datasets and missingness patterns.

Benchmark Table

The table below is rendered with DataTables, the same third-party display library used in How to Use.

Methodology

Benchmarks are computed by injecting synthetic missingness into each dataset, imputing, and scoring only the masked entries against the ground truth. Two missingness patterns are evaluated:

  • MAR_0.10: 10% missing-at-random across all cells.

  • Blocks_0.20x0.30: contiguous blocks covering 20% of the rows in 30% of the columns.

Two dataset families are covered (see the family and imputer columns):

  • Tabular datasets are imputed with MultivariateImputer. Rows with pre-existing missing values are dropped first so every masked cell has a ground truth. Datasets include numeric-only sets widely used in the imputation literature (Diabetes, Wine, Breast Cancer, California Housing, Wine Quality red, Spambase) and mixed numeric/categorical sets (Letter Recognition, Abalone, Ionosphere, Titanic, plus a synthetic mixed dataset).

  • Time series datasets are imputed with TimeSeriesImputer (lags/leads 1-3, default time features). The time grid is preserved: rows are never dropped, and synthetic missingness is only injected into observed cells, so pre-existing gaps stay untouched and every masked cell can be scored. Datasets are standard time series imputation benchmarks: the PEMS-BAY and METR-LA traffic-speed datasets (first four weeks of 5-minute data, first 60 sensors, imputed with n_nearest_features=100; zero readings in METR-LA are treated as missing), the Beijing PM2.5 hourly air-quality dataset (mixed: numeric measurements plus the categorical wind direction), and the ETTh1 electricity-transformer dataset.

All external datasets are downloaded and cached with pooch from their canonical open sources (UCI Machine Learning Repository, Zenodo, the ETDataset repository) and verified against pinned checksums.

Metrics are split by data type: regression metrics (RMSE, MAE, R2, MAPE, SMAPE, median AE, bias, normalized RMSE) for numeric columns and classification metrics (accuracy, balanced accuracy, macro precision/recall/F1, MCC, Cohen’s kappa) for categorical columns. Coverage reports the fraction of masked values that received finite predictions.