Introduction
Deep learning models are inherently data-intensive, yet in critical domains such as medicine, finance, and astronomy, labeled time series data is often scarce, expensive, or unbalanced. Traditional data augmentation techniques like time slicing or time warping frequently fail to preserve long-term temporal correlations or struggle with irregular sampling—where data points are recorded at inconsistent intervals. This paper introduces T-CGAN (Time-Conditional Generative Adversarial Network), a framework designed to generate realistic synthetic sequences from noisy and irregularly sampled training data.
Methodology
The core of T-CGAN is a variation of the Conditional Generative Adversarial Network (CGAN) that explicitly incorporates temporal information.
Architecture
The model consists of two primary components:
- Generator ($G$): Maps random noise ($z$) and a vector of timestamps ($t$) to a synthetic time series ($g$). It utilizes deconvolutional layers to ensure the output matches the provided temporal coordinates.
- Discriminator ($D$): Evaluates pairs of time series and timestamps. Using convolutional layers, it learns to distinguish between real data and synthetic samples by assessing the realism of values relative to their specific recording times.
By training on (value, timestamp) pairs, T-CGAN treats time as a continuous coordinate rather than a fixed index, allowing it to "fill in the gaps" of irregularly sampled data.
Experimental Evaluation
The authors evaluated T-CGAN across synthetic and real-world benchmarks, including Starlight Curves, Power Demand data, and ECG200 heart activity records. To test robustness, datasets were intentionally unbalanced and subjected to random data point deletions (10% to 40% missingness).
Visual Fidelity
A key strength of T-CGAN is its ability to preserve the global structure of the data. Unlike local window-based methods, T-CGAN generates entire sequences, maintaining long-term correlations. Visual inspections of synthetic sine and sawtooth waves demonstrate that the model accurately captures the underlying dynamics of noisy, irregularly sampled signals.
Results and Discussion
T-CGAN consistently outperformed traditional methods like Time Slicing and Time Warping. On the Starlight dataset, T-CGAN augmentation improved the AUROC from 0.71 to 0.98. Furthermore, the model demonstrated high robustness to missing data, maintaining an AUROC of 0.975 even with 20% of data points removed. By generating synthetic samples for minority classes, T-CGAN effectively re-balances datasets, leading to more reliable classification performance.
Conclusion
T-CGAN provides a robust solution for augmenting "messy" time series data. Future work may explore scaling the model to high-frequency, long-duration sequences and extending the framework to multivariate time series.
Summary updated