Work

TDAssure: a topological data analysis framework for assessing missing data imputation quality

Year: 2026
Type: preprint
Abstract: Conventional measures of imputation quality assess how accurately missing values are reconstructed but provide limited information about preservation of the multivariate structure of the data. We introduce TDAssure, a topological data analysis framework for assessing structural preservation following missing-data imputation. The framework constructs Mapper graphs for complete and imputed cases, summarises their topological features using extended persistence diagrams, and quantifies differences using bottleneck distances. Permutation tests assess the statistical significance of these differences across a predefined Mapper parameter grid, with adjustment for multiple testing. We evaluated the framework on the Fashion-MNIST and Epileptic Seizure Recognition datasets using four imputation methods under missing completely at random, missing at random, and missing not at random mechanisms, with target missingness rates of 20% and 30%. Random imputation consistently produced the greatest reconstruction error and the largest number of Mapper configurations showing significant topological differences. The relative performance of k-nearest-neighbour, autoencoder, and mean imputation varied between datasets and assessment approaches. In particular, lower reconstruction error did not always correspond to fewer detectable topological differences, with disagreements most apparent when data were missing not at random. Interpretation in such settings requires caution because complete cases may provide an unrepresentative structural reference. An interactive. TDAssure, implemented in a Shinny application, provides a complementary assessment of imputation quality by examining structural preservation alongside pointwise reconstruction accuracy.
Language: English
Open Access status: green
OpenAlex ID: W7235495659
Cites: 24
Cited by:
Related to: 10