Time-Series GAN (TimeGAN)

Maps the high-dimensional time-series features into a lower-dimensional latent space that captures the underlying temporal structure. This adversarial latent encoding allows the generator to learn the distribution of sequences rather than just individual points, enabling the model to handle complex, non-linear dynamics present in real-world data like patient vitals or stock prices.

Performs the inverse mapping from the latent representation back to the original feature space. This component ensures that the latent dynamics are invertible and that the synthetic sequences can be reconstructed without losing critical statistical properties. The recovery loss acts as a regularizer, preventing the generator from collapsing to trivial solutions.

A recurrent neural network operating in the latent space that generates synthetic latent codes autoregressively. Unlike standard GANs that generate independent samples, this generator learns the conditional distribution P(Z_t | Z_{t-1}), capturing the stepwise temporal dependencies that define realistic sequences. It receives both random noise and supervised feedback from the discriminator.

Operates on the latent codes rather than raw features, distinguishing between real and synthetic latent sequences. This design choice is critical: by discriminating in the learned representation space, the model focuses on temporal coherence rather than just pointwise realism. The discriminator provides the adversarial loss that drives the generator toward producing indistinguishable sequences.

A unique element of TimeGAN that captures stepwise conditional distributions using a teacher-forcing objective. While the adversarial loss ensures the overall sequence distribution matches real data, the supervised loss explicitly trains the generator to predict the next latent vector given the previous one. This hybrid objective prevents mode collapse and ensures temporal consistency.

Combines three loss terms into a unified optimization:

• Adversarial loss: Minimax game between generator and discriminator
• Supervised loss: Maximum likelihood of next-step transitions
• Reconstruction loss: Fidelity of embedding-recovery cycle

This tripartite objective enables TimeGAN to outperform purely adversarial or purely autoregressive models on both discriminative and predictive downstream tasks.

The key innovation distinguishing TimeGAN from standard GANs. Beyond the unsupervised adversarial loss, TimeGAN incorporates a stepwise supervised loss using the real data's temporal transitions. This explicitly trains the model to capture the conditional distribution P(X_t | X_{t-1}), preventing the generator from ignoring the sequential structure and ensuring temporal coherence in generated sequences.

A quantitative fidelity metric for time-series. A post-hoc classifier is trained to distinguish real sequences from generated ones. The discriminative score is the classification error; a score near 0.5 indicates the synthetic data is indistinguishable from real data. This complements visual inspection by providing a rigorous statistical test of distributional similarity.

A critical privacy audit for TimeGAN models. An adversary attempts to determine if a specific patient's record was part of the GAN's training set by analyzing the generator's outputs. A model vulnerable to this attack leaks training data information. Robust TimeGAN implementations must be hardened against such attacks to ensure patient confidentiality.