A Non-Stationary Synthetic Data Generator for Longitudinal Data

Jhajharia, Prasun ; Grimm, Mirko

4. term

Computer Science (IT), Master

2025

2025-06-13

21 pages

Longitudinal datasets, which capture repeated observations of individuals over time, are important in areas such as healthcare, finance, and education. However, the right to privacy as stated in the General Data Protection Regulation (GDPR) limits the direct sharing of such sensitive data. Syn- thetic data generation offers a key solution by producing synthetic datasets that captures key statistical characteris- tics of real data while reducing the risk of exposing sensitive information. While models like Bayesian networks are use- ful for generating synthetic data and are easy to interpret, they’re mainly built for static datasets. This means they often fall short when it comes to handling time based patterns. In this work, we propose a framework for generating synthetic longitudinal categorical data using segment-wise Dynamic Bayesian Networks (DBNs). Our method detects change points in the temporal data to identify non-stationary segments and learns separate DBNs for each segment. These models are then aggregated by extracting common struc- tural patterns, and segment wise behaviors are grouped via clustering. Synthetic sequences are generated by sampling from these clusters, preserving both temporal coherence and structural realism. We tested our approach on both simulated and real-world datasets, including Electronic Health Records (EHRs). While our model shows lower utility compared to our baselines, it offers better privacy protection demonstrat- ing that a modest trade-off in accuracy can lead to significant gains in protecting sensitive information.

Download
View record in AAU Student Projects