Institutional Repository of Vilnius University
Sintetinių duomenų generavimo metodų palyginimas /
Abstract
dc:descriptionThis thesis examines synthetic data generation methods and their suitability for ensuring the confidentiality of service sector company data. The aim of the study is to evaluate the ability of different synthetic data generation methods to preserve analytical utility while reducing disclosure risk. The analysis was conducted using 2025 service sector enterprise activity survey (PS20) data, including monthly revenue, annual revenue, number of employees, and categorical company characteristics. Three synthetic data generation methods were applied in this study: the parametric Generalized Additive Models for Location, Scale, and Shape (GAMLSS) method, the non-parametric The Fully Conditional Specification (FCS) method with the Classification and regression trees (CART) algorithm, and the deep learning-based Conditional Tabular Generative Adversarial Network (CTGAN) method. The GAMLSS method was applied for modelling monthly revenue variables, while the FCS and CTGAN methods were used to generate complete synthetic datasets. In order to reduce data asymmetry and the influence of extreme values, a sign-preserving logarithmic transformation was applied in part of the models. The quality of synthetic data was evaluated by comparing the distributions of original and synthetic data using the Kolmogorov--Smirnov test and graphic density analysis. Disclosure risk was assessed using the UiO, repU, Dorig, and DiSCO indicators. The results showed that the FCS method with the CART algorithm achieved the best balance between data quality and confidentiality protection. This method most accurately reproduced the distributions of monthly revenue variables, and the Kolmogorov--Smirnov test did not indicate statistically significant differences between the original and synthetic data distributions. In addition, significantly lower identity and attribute disclosure risks were observed in the synthetic data. The results of the GAMLSS method showed that the sign-preserving logarithmic transformation improved the similarity between synthetic and original data distributions, although statistically significant differences remained. The CTGAN method ensured low disclosure risk but did not accurately reproduce the distributions of monthly revenue variables. The results of the study suggest that the FCS method is the most suitable for the analysed service sector dataset.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Stankevičiūtė, Simona,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308333313&prefLang=en_US