Diffusion Transformers for Tabular Data Time Series Generation
Fabrizio Garuti, Enver Sangineto, Simone Luetto, Lorenzo Forni, Rita Cucchiara
TL;DR
This work addresses the generation of time series composed of heterogeneous tabular rows with variable length. It introduces TabDiT, a latent diffusion model built on a Diffusion Transformer that encodes individual rows with an autoregressive VAE and then uses a Transformer denoiser to assemble temporally coherent sequences, while handling variable lengths via end-of-sequence signaling. A variable-range numeric representation and an autoregressive VAE decoder enable robust encoding/decoding of mixed numerical and categorical features. Across six public datasets and a large-scale bank dataset, TabDiT outperforms strong baselines in both unconditional and conditional settings, achieving strong diversity and realism and demonstrating scalability to long sequences and large data volumes.
Abstract
Tabular data generation has recently attracted a growing interest due to its different application scenarios. However, generating time series of tabular data, where each element of the series depends on the others, remains a largely unexplored domain. This gap is probably due to the difficulty of jointly solving different problems, the main of which are the heterogeneity of tabular data (a problem common to non-time-dependent approaches) and the variable length of a time series. In this paper, we propose a Diffusion Transformers (DiTs) based approach for tabular data series generation. Inspired by the recent success of DiTs in image and video generation, we extend this framework to deal with heterogeneous data and variable-length sequences. Using extensive experiments on six datasets, we show that the proposed approach outperforms previous work by a large margin.