Missing data are a ubiquitous challenge when training unbiased, high confidence machine learning (ML) models on real-world data. To help address this issue, we propose SudokuImputer v1.0.0, a novel graph-based framework, to estimate missing values in an iterative, uncertainty aware process. The behavior of SudokuImputer is assessed across numerous design hyperparameters, including multiple network centrality methods, feature prioritization modes, statistical associations and pairwise availability ratios for edge weight assignment, and partner node proportions. We benchmark SudokuImputer against point-value imputations, MICE, kNN, matrix factorization, and SoftImpute. SudokuImputer achieves best-in-class RMSE on MAR (mean rank = 1.7) and MNAR (mean rank = 1.8) missing data across most missingness proportions from ten percent to fifty percent in three experimental benchmarks. SudokuImputer is sensitive to dataset dimensionality, and optimizing algorithmic runtime remains an unresolved challenge. Future work should evaluate SudokuImputer on mixed-data benchmarks and seek to iterate on the foundational graph framework laid here.
