Abstract
Multivariate time series (MTS) data, such as biometric signals from a single patient, are vital for research but often restricted by privacy regulations. Synthetic data generation via Generative Adversarial Networks (GANs) is a promising alternative; however, maintaining realistic inter-channel correlations remains a significant challenge. We present COSCI-GAN (COmmon Source CoordInated GAN), a novel framework designed to generate high-fidelity MTS. By utilizing a shared latent space and a Central Discriminator, COSCI-GAN successfully captures both individual channel dynamics and complex cross-channel dependencies. Our experiments demonstrate that COSCI-GAN outperforms existing state-of-the-art methods in preserving data utility and correlation structures.
Introduction
In domains like medicine and finance, multivariate data—where multiple variables are recorded over time—is highly sensitive. Sharing a patient's concurrent heart rate, blood pressure, and oxygen levels is difficult due to privacy laws. While GANs can generate synthetic versions of this data, they face a dilemma: learning the joint distribution of all channels simultaneously is computationally prohibitive, while generating channels independently destroys the critical correlations between them. If synthetic signals do not "sync up" realistically, the data becomes useless for training diagnostic models.
Methodology
The COSCI-GAN architecture addresses this by introducing a coordinated approach to generation. The framework rests on two primary pillars:
- Shared Latent "Patient Space": Instead of independent noise sources, all channel generators are fed the same noise vector $z$. This vector represents the "common source" or the overall environment (e.g., a specific patient's physiological state) from which all signals originate.
- Dual Objective Optimization:
- Channel GANs (Local): Each data channel is assigned a dedicated Generator-Discriminator pair. These local units focus on mastering the specific temporal patterns and "rhythms" of their respective channels.
- Central Discriminator (Central): A single, global discriminator evaluates the entire set of generated channels as a unified instance. Its role is to ensure that the inter-channel relationships in the synthetic data match those found in the real dataset.
The model balances these local and central objectives using a hyperparameter $\gamma$, allowing it to produce high-quality individual signals that are also properly synchronized.
Experimental Setup
COSCI-GAN was evaluated against state-of-the-art (SOTA) models, including TimeGAN and Fourier Flows, across several datasets:
- Toy Sine Datasets: Simulated datasets with known, controllable correlations (e.g., frequency changes and anomalies).
- EEG Eye State Dataset: A 14-channel real-world dataset used for classifying eye states based on brain activity.
- Google Stocks Dataset: Historical stock data used to assess the diversity of generated samples.
Key Results
COSCI-GAN consistently outperformed SOTA benchmarks in both statistical similarity and practical utility.
Correlation Preservation
In the EEG dataset, COSCI-GAN achieved a Mean Absolute Error (MAE) for correlation similarity of 0.111, a marked improvement over TimeGAN (0.257) and Fourier Flows (0.146). Visual analysis via heatmaps confirms that the Central Discriminator (CD) is essential for recovering the correlation structure of the real data.
Downstream Utility
The ultimate test of synthetic data is its performance in training models. When a classifier was trained on synthetic data and tested on real data, the COSCI-GAN version achieved an accuracy of 0.634, compared to 0.561 for a version lacking the Central Discriminator. Furthermore, PCA and t-SNE visualizations indicated that COSCI-GAN’s synthetic samples had significantly higher overlap with real data distributions than competing methods.
Discussion and Insights
The research highlights a fundamental correlation-diversity trade-off: enforcing strict inter-channel correlations can sometimes slightly degrade the accuracy of individual channel distributions. Finding the optimal balance is crucial for data utility.
Interestingly, the authors found that while LSTM-based networks were ideal for individual channel generators, a simpler Multi-Layer Perceptron (MLP) often performed better for the Central Discriminator. A discriminator that is too complex can hinder the generators, causing them to sacrifice signal quality to satisfy correlation constraints. Additionally, the modular nature of COSCI-GAN allows for parallel training of individual channel GANs, offering potential scalability benefits.
Conclusion and Future Work
While COSCI-GAN represents a significant step forward, challenges remain. Scaling the model to thousands of channels may require substantial computational resources. Furthermore, while synthetic data offers inherent privacy, the model does not yet provide formal mathematical guarantees like Differential Privacy (DP). Future work will focus on integrating DP into the training loop and exploring more complex multi-source latent embeddings.
Summary updated