Abstract
Large-scale high-quality data is critical for training modern deep neural networks. However, data acquisition can be costly or time-consuming for many time-series applications, thus researchers turn to generative models for generating synthetic time-series data. In particular, recent generative adversarial networks (GANs) have achieved remarkable success in time-series generation. Despite their success, existing GAN models typically generate the sequences in an auto-regressive manner, and we empirically observe that they suffer from severe distribution shifts and bias amplification, especially when generating long sequences. To resolve this problem, we propose Adversarial Error Correction GAN (AEC-GAN), which is capable of dynamically correcting the bias in the past generated data to alleviate the risk of distribution shifts and thus can generate high-quality long sequences. AEC-GAN contains two main innovations: (1) We develop an error correction module to mitigate the bias. In the training phase, we adversarially perturb the realistic time-series data and then optimize this module to reconstruct the original data. In the generation phase, this module can act as an efficient regulator to detect and mitigate the bias. (2) We propose an augmentation method to facilitate GAN's training by introducing adversarial examples. Thus, AEC-GAN can generate high-quality sequences of arbitrary lengths, and the synthetic data can be readily applied to downstream tasks to boost their performance. We conduct extensive experiments on six widely used datasets and three state-of-the-art time-series forecasting models to evaluate the quality of our synthetic time-series data in different lengths and downstream tasks. Both the qualitative and quantitative experimental results demonstrate the superior performance of AEC-GAN over other deep generative models for time-series generation.
Introduction
Generating long, realistic time-series data is a significant challenge in machine learning. Most current state-of-the-art generators, such as TimeGAN or SigCWGAN, utilize an auto-regressive approach where each new data point is predicted based on a window of previously generated points. While effective for short-term generation, this method is prone to bias amplification and distribution shift. Much like the game of "Telephone," small numerical errors at early steps compound over time, leading to significant distortions in long-horizon sequences. These distortions often manifest as "dying" amplitudes or unstable frequencies, rendering the synthetic data less useful for training downstream forecasting models.
The AEC-GAN Framework
To address these issues, we introduce the Adversarial Error Correction GAN (AEC-GAN). The core philosophy is to treat the previously generated history not as ground truth, but as a potentially noisy signal that requires refinement.
Adversarial Error Correction Module ($M_\omega$)
AEC-GAN incorporates a dedicated correction module that dynamically "denoises" the generated history before it is used to predict the next step. This module is trained using a "perturb-and-reconstruct" strategy. During training, real time-series data is subjected to adversarial noise (via PGD-attacks) to simulate typical generation errors. The module is then optimized to reconstruct the original, clean signal. During the inference/generation phase, this module acts as a regulator, detecting and mitigating bias in real-time.
Adversarial Augmentation
To further enhance robustness, the generator is trained using both clean and adversarially perturbed conditions. This adversarial training forces the model to learn a more stable mapping, ensuring it can produce high-quality outputs even when the input history contains slight inaccuracies. Architecturally, the framework utilizes a 1D-convolutional neural network (Conv1d) for the error correction module, while the generator and discriminator are built using an auto-regressive feed-forward network (AR-FNN) architecture.
Experimental Evaluation
We evaluated AEC-GAN across six diverse datasets representing different domains:
- Energy: ETTh1, ETTh2, ETTm1, ETTm2 (Electricity Transformer Temperature).
- Population: US Births.
- Disease: ILI (Influenza-Like Illness).
These datasets provide a range of frequencies and sequence lengths, testing the model's ability to maintain long-term stability.
Results and Discussion
Quantitative analysis using the Frechet Distance (FD) score reveals that AEC-GAN consistently outperforms existing models. As sequence lengths increase (up to 4000 steps), the FD scores for models like TimeGAN often explode, indicating a breakdown in data quality. In contrast, AEC-GAN maintains a stable and low FD score.
Qualitatively, the visual fidelity of AEC-GAN is superior. As shown in the figure above, while SigCWGAN suffers from significant amplitude attenuation (the signal flattens over time) and TimeGAN exhibits frequency instability, AEC-GAN successfully preserves the amplitude and periodic structure of the real data.
Furthermore, synthetic data generated by AEC-GAN significantly improves downstream task performance. When training a forecasting model (SCINet) on synthetic data, AEC-GAN achieved a 17.9% improvement on average compared to other generative benchmarks, demonstrating its utility for data augmentation in real-world applications.
Conclusion
AEC-GAN effectively mitigates the distribution shift inherent in auto-regressive time-series generation through its adversarial error correction module and augmentation strategy. By stabilizing long-term generation, it provides a robust tool for creating high-quality synthetic datasets that can enhance the performance of downstream forecasting models.
Summary updated