Mask-Aware Smoothness-Regularized GANs for Spatio-Temporal Data Imputation

Spatio-temporal data, especially in biomedical applications, often exhibit structurally complex and high rates of missingness, which can simultaneously distort both the marginal distribution of the data and its spatio-temporal dependence structure. Generative adversarial networks (GANs), as a powerful class of generative models, do not explicitly model or estimate the data density; instead, they learn, via adversarial training, a generator that produces samples that are difficult to distinguish from real data, thereby implicitly approximating the target distribution. Such an implicit modeling mechanism is more flexible and expressive when dealing with complex spatio-temporal distributions induced by missingness patterns. Existing generative imputation methods based on MISGAN can, in principle, recover the observed-data distribution, but they do not impose explicit spatio-temporal smoothness; in contrast, existing spatio-temporal adversarial imputation models typically encode spatio-temporal dependence implicitly through the network architecture, and they tend to use the missingness mask in a relatively coarse manner, making it difficult to systematically characterize the underlying missing-data mechanism. Building on the MISGAN framework, this paper proposes Spatio-Temporal Regularized MISGAN (ST-MISGAN), a mask-aware, spatio-temporally regularized adversarial imputation method, and further introduces a novel imputation procedure that jointly models the observed data and the missingness mask, so as to balance distributional recovery accuracy and downstream task performance under complex missing-data mechanisms.