Diffusion Transformers for Tabular Data Time Series Generation
Authors: Fabrizio Garuti, Enver Sangineto, Simone Luetto, Lorenzo Forni, Rita Cucchiara
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Using extensive experiments on six datasets, we show that the proposed approach outperforms previous work by a large margin. Due to the lack of a unified evaluation protocol for tabular data time series generation, we collect different public datasets and we propose an evaluation metric which extends single-row metrics to the time-series domain. For the specific case of unconditional generation, Tab Di T is, to the best of our knowledge, the first deep learning method for unconditional generation of heterogeneous tabular data time series. In this case, we compare Tab Di T with a strong AR baseline, which we implemented by merging the (discriminative) hierarchical architecture proposed in (Padhi et al., 2021; Luetto et al., 2023), with some of the architectural solutions we propose in this paper. In all the experiments, Tab Di T significantly outperforms all the compared baselines, usually by a large margin. Section 5 is dedicated to 'EXPERIMENTS', including 'EVALUATION PROTOCOL', 'ABLATION' and 'MAIN EXPERIMENTS' sections. |
| Researcher Affiliation | Collaboration | Fabrizio Garuti1,2, Enver Sangineto2, Simone Luetto3, Lorenzo Forni1 & Rita Cucchiara2 1 Prometeia Associazione, Bologna, Italy, 2 University of Modena and Reggio Emilia, Italy, 3 Prometeia Sp A, Bologna, Italy. EMAIL, * EMAIL. The authors are affiliated with both 'Prometeia Associazione' and 'Prometeia Sp A' (industry) and 'University of Modena and Reggio Emilia' (academia). |
| Pseudocode | No | The paper describes methods and algorithms using mathematical formulations and textual descriptions. It does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | The evaluation protocol code is available at: https://github.com/fabriziogaruti/TabDiT |
| Open Datasets | Yes | We use six public datasets, whose statistics are provided in App. E. Age1, Age2, Leaving, taken from (Fursov et al., 2021), and PKDD 99 Financial Dataset, taken from (Berka, 1999), are composed of bank transaction time series of different real banks with different attributes. On the other hand, the Rossmann and the Airbnb datasets, used in (Patki et al., 2016; Solatorio & Dupriez, 2023), are composed of, respectively, historic sales data for different stores and access log data from Airbnb users. Original source, copyright, and license information are available in the links in the footnotes. |
| Dataset Splits | Yes | Following the protocol adopted in (Solatorio & Dupriez, 2023), all the experiments of this section have been repeated three times with different random splits of the samples between the generator training data, the discriminators training data and the testing data (see App. C and E for more details). In Rossmann, we use 80% of the store data and their associated sales records for training the generator. We use the remaining stores as the testing data. In the Airbnb dataset, we consider a random sample of 10,000 users for the experiment. We take 8,000 as part of our training data, and we assess the metrics using the 2,000 users in the testing data. Table 9: Dataset statistics. (This table provides 'Training samples' and 'Testing samples' counts for each dataset). |
| Hardware Specification | Yes | All the experiments presented in this paper have been performed on an internal compute node composed of: 2 CPUs AMD EPYC 7282 16-Core, for a total of 32 physical and 64 logical cores, 256 Gb RAM, 4 GPUs Nvidia RTX A6000, each with 48 Gb of memory each, for a total of 192 Gb. |
| Software Dependencies | No | The paper mentions using 'Adam W' as the optimizer (Table 11, Table 12) and 'Cat Boost' (Sec 5.1), but does not provide specific version numbers for these or any other software libraries or programming languages used. |
| Experiment Setup | Yes | Table 10: Dataset-specific hyperparameter values. (Includes VAE Training epochs, Training iterations, Denoising network Training epochs, Training iterations). Table 11: Dataset-independent hyperparameter values for the VAE. (Includes Optimizer, Learning rate, Training dropout, Batch size, Model size, VAE Encoder Transformer layers/heads/hidden size, VAE Decoder Transformer layers/heads/hidden size, VAE latent space size, βmax, βmin, λ, patience). Table 12: Dataset-independent hyperparameter values of the denoising network. (Includes Optimizer, Positional encoding, Learning rate, Dropout, Batch size, Model size, Di T depth, Di T num heads, Hidden size (d), Diffusion steps, pd, s). |