Temporal-Aware Synthetic Data Generation in Multi-Relational Tabular Data

Abstract

Multi-relational time-series tabular data is widely used in fields such as e-commerce, finance, and healthcare. Recent advances in synthetic tabular data generation techniques have improved statistical similarity and relational fidelity, but previous methods struggle to preserve temporal transition patterns within child sequences. This paper proposes a framework that improves the temporal fidelity of multi-relational time-series tabular data synthesis by integrating a Temporal Representation Encoder (TRE) and a diffusion-based generative model. TRE uses a transformer to learn embeddings that capture both within-row field interactions and temporal dependencies in sequence through masked language modeling. By performing clustering in this temporal-aware embedding space, rather than in the raw feature space, we obtain cluster labels that meaningfully group sequences with similar temporal patterns. These cluster labels can be easily integrated into various diffusion models. Furthermore, we introduce new metrics for temporal evaluation, including comparisons of transition matrices using L1 distance and Jensen-Shannon divergence, and lag-k difference analysis of numerical features overlooked in previous work. We evaluate our approach by integrating TRE with two state-of-the-art diffusion models, ClavaDDPM and TabDiT. We conduct comprehensive experiments on two toy examples and three real-world datasets (Rossmann, Airbnb, and Walmart). On the Rossmann dataset, ClavaDDPM + TRE captures both short-term differences and weekly patterns better than the vanilla model. Similar improvements are observed across datasets. These findings demonstrate that clustering within the learned temporal embedding space provides a more effective conditioning mechanism for preserving sequential dynamics in synthetic multi-relational time-series tabular data, suggesting new directions for generating time-aware synthetic data.