Aproximace agregovaných dat na nižší úrovně

Disaggregating socio-economic indicators from coarse regional aggregates to finer spatial units is a recurring challenge in public policy and statistical analysis. While traditional methods such as areal interpolation and regression assume full supervision or spatial continuity, they fall short when fine-grained ground truth is unavailable. This thesis proposes a modular, simulation-based pipeline for approximating fine-resolution values from coarse inputs using auxiliary features. The method combines unsupervised clustering, simulation of district-level targets consistent with known regional aggregates, and supervised learning on the synthetic data. It is designed to generalize across different indicators and domains. The pipeline is evaluated on two tasks: unemployment prediction and education level estimation at the district level in the Czech Republic. Using voting data and household structure as auxiliary inputs, the method consistently outperforms strong baselines and reproduces intra-regional variation more faithfully than models trained directly on coarse labels. The results demonstrate that meaningful disaggregation is possible even under partial supervision, offering a scalable solution for real-world deployment in data-scarce environments.