Generative Data Imputation with Uncertainty Quantification for Evaluation of Downstream Analysis Sensitivity

Escudero-Arnanz et al. (2024) show that sophisticated tensor completion can outperform traditional imputation, yet most downstream analyses treat imputed values as “real,” ignoring uncertainty. This idea proposes developing a generative imputation framework that outputs distributions over possible values, not just point estimates. By integrating these uncertainty distributions into subsequent analyses (classification, regression, mediation, etc.), researchers can assess how sensitive their conclusions are to different missing data scenarios—addressing concerns raised by Gabr et al. (2023) about performance evaluation under missingness. The innovation is in coupling state-of-the-art generative models (building on the GAN-VAE work in Sekhar et al., 2025) with robust propagation of uncertainty, enabling more honest and transparent inference in fields where missingness is pervasive. Impact: More reliable science, policy, and clinical decision-making under uncertainty.

References:

1. Low-Rank Tensor Completion for Heart Failure Exacerbation Detection in Multivariate Time Series with Missing Data. Óscar Escudero-Arnanz, Rosa Sicilia, Cristina Soguero-Ruíz, Inmaculada Mora-Jiménez, Diana Lelli, Claudio Pedone, Antonio G. Marques (2024). 2024 IEEE 37th International Symposium on Computer-Based Medical Systems (CBMS).
2. Effect of Missing Data Types and Imputation Methods on Supervised Classifiers: An Evaluation Study. Menna Ibrahim Gabr, Y. Helmy, D. Elzanfaly (2023). Big Data and Cognitive Computing.
3. EV Charging Station Management using Model Predictive Control and Generative Adversarial Variational Autoencoders. P. C. Sekhar, Harinath. G, Mujahid K, Pavani D (2025). 2025 3rd International Conference on Data Science and Information System (ICDSIS).