Sequential RC-TGAN for Relational Time Series

• The paper introduces a GAN framework that extends RC-TGAN by using a conditional recurrent generator to synthesize entire time-ordered sequences conditioned on static parent attributes.
• It incorporates a differentiable spectral envelope loss—augmented by Variational Gaussian Mixture discretization for continuous features—to enforce latent cyclic and seasonal patterns.
• Empirical results show that Seq. RC-TGAN outperforms baselines on spectral, autocorrelation, and envelope divergence metrics across both simulated and real relational datasets.

Searching arXiv for Sequential RC-TGAN and closely related work to ground the article in the cited literature. First, I’ll look up the Seq. RC-TGAN paper itself and then a few directly relevant related papers. Sequential RC-TGAN (Seq. RC-TGAN) is a GAN-based framework for synthetic multi-table relational databases in which the child tables contain time series. It extends RC-TGAN from static relational row generation to conditional sequence generation, and augments the adversarial objective with a differentiable loss derived from Spectral Envelope Theory in order to preserve latent periodic structures such as seasonality and cycles in categorical and continuous features. The framework is defined on a parent table $W$ of static entities and a child table $U$ of time-ordered rows, with the modeling target given by the conditional joint distribution $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$ (Gueye et al., 30 Jun 2026).

1. Relational-temporal problem formulation

Seq. RC-TGAN is designed for schemas $\mathcal{S}=\{W,U\}$ in which a parent row $w\in W$ is associated with an ordered child sequence $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$, where each $u_t\in\mathcal{V}(U)$. The parent table contains static attributes $w.\mathcal{A}$, while the child table contains multivariate temporal rows with categorical attributes $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$ and continuous attributes $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$. Within this formulation, the objective is not merely to synthesize plausible rows, but to preserve relational structure, short- and long-range temporal dependencies, and latent periodic structures across the full child trajectory (Gueye et al., 30 Jun 2026).

The central extension over RC-TGAN consists of two changes. First, the static generator is replaced by a conditional recurrent generator that emits a whole sequence $U$0 conditioned on the parent row. Second, the generator is regularized by a differentiable spectral envelope loss $U$1, introduced specifically because one-hot encodings of categorical sequences do not by themselves capture cyclic or seasonal structure in the frequency domain. The paper further extends this frequency-domain regularization to continuous features through Variational Gaussian Mixture (VGM) discretization, thereby providing a unified treatment of mixed-type relational time series (Gueye et al., 30 Jun 2026).

Formally, Seq. RC-TGAN learns a conditional generator $U$2 such that

$U$3

where $U$4 are i.i.d. noise vectors. The corresponding factorization

$U$5

makes explicit that relational conditioning and temporal dependence are modeled jointly rather than as separate post hoc constraints (Gueye et al., 30 Jun 2026).

2. Generator and critic architecture

The generator replaces the original RC-TGAN MLP row generator with a recurrent network, stated in the paper as “e.g., RNN/LSTM.” At each time step,

$U$6

with $U$7 initialized, for example, at zeros, and $U$8 denoting concatenation. The conditioning variable $U$9 is concatenated at every time step, so the parent context is injected throughout the sequence rather than only at initialization (Gueye et al., 30 Jun 2026).

This design has three stated properties. Conditioning on the parent at every step preserves relational context across the entire sequence. The auto-regressive hidden state $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$0 captures temporal dependencies. The MLP output head is mixed-type: it produces logits for categorical variables and parameters or direct values for continuous variables in a standard tabular GAN style. A plausible implication is that the model uses the recurrent hidden state as the principal carrier of temporal structure, while the per-step noise provides local stochasticity conditioned on parent identity.

The critic is a conditional MLP operating on the flattened child sequence concatenated with parent attributes: $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$1 Its output is a scalar score in a WGAN formulation, with higher scores assigned to real than to generated sequences. The critic therefore evaluates global coherence of the full sequence under relational conditioning, rather than scoring individual time steps independently. This is a notable architectural choice because it places adversarial pressure on whole-trajectory realism rather than on local transitions alone (Gueye et al., 30 Jun 2026).

3. Spectral envelope theory and the loss construction

The distinctive component of Seq. RC-TGAN is its use of Spectral Envelope Theory as a differentiable generator-side regularizer. For a stationary real-valued process $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$2, spectral density is defined by

$\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$3

with $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$4. For a length-$\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$5 sequence, the periodogram at Fourier frequencies $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$6 is

$\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$7

For multivariate processes, the scalar spectral density is replaced by a spectral density matrix and the periodogram by $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$8 (Gueye et al., 30 Jun 2026).

Categorical time series require a different construction. Let $\mathbb{P}(u_1,\dots,u_T \mid w.\mathcal{A})$9 have state space $\mathcal{S}=\{W,U\}$0. Spectral envelope theory considers all numeric scalings $\mathcal{S}=\{W,U\}$1 of the categories and defines

$\mathcal{S}=\{W,U\}$2

Using the one-hot encoded process $\mathcal{S}=\{W,U\}$3, with spectral density matrix $\mathcal{S}=\{W,U\}$4 and covariance matrix $\mathcal{S}=\{W,U\}$5, this becomes

$\mathcal{S}=\{W,U\}$6

The solution is the largest generalized eigenvalue of $\mathcal{S}=\{W,U\}$7, and the maximizing vector $\mathcal{S}=\{W,U\}$8 is the optimal scaling at frequency $\mathcal{S}=\{W,U\}$9 (Gueye et al., 30 Jun 2026).

The paper places spectral envelopes in $w\in W$0, which permits the loss

$w\in W$1

For a mini-batch $w\in W$2 and a categorical feature $w\in W$3, the batch-averaged real and synthetic envelopes are computed and compared through

$w\in W$4

Because the computation is expressed through matrix operations and eigenvalue decompositions on the synthetic batch, gradients $w\in W$5 can be backpropagated to the generator. The intended effect is to align peak locations and amplitudes in the spectrum, preventing generated sequences from matching time-domain realism while remaining frequency-domain deficient (Gueye et al., 30 Jun 2026).

A common misconception is that one-hot encoded categorical sequences already preserve the periodic structure needed for frequency matching. The Seq. RC-TGAN formulation explicitly rejects that assumption: the spectral envelope is introduced precisely because one-hot encoding does not expose the latent cyclic structure in a way that standard losses can exploit.

4. Extension to continuous attributes through VGM discretization

Spectral envelope theory is categorical by construction, so Seq. RC-TGAN extends it to continuous features through a Variational Gaussian Mixture Model. For each continuous attribute $w\in W$6, the real data are fit with a VGM having $w\in W$7 components: $w\in W$8 The most likely mode is then assigned by

$w\in W$9

yielding a discrete mode sequence $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$0 suitable for spectral envelope computation (Gueye et al., 30 Jun 2026).

To retain within-mode information, each continuous value is also represented by a normalized intra-mode scalar,

$\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$1

The full representation of a continuous observation therefore comprises a one-hot mode indicator and the scalar $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$2. The paper states that this representation is fully invertible given learned $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$3, so discretization is used only for spectral envelope computation rather than as an information-destroying replacement of the original value (Gueye et al., 30 Jun 2026).

The continuous spectral loss is defined analogously to the categorical case: $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$4 The overall spectral penalty is

$\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$5

This weighting gives a unified frequency-domain objective across mixed data types. A plausible implication is that the model’s notion of temporal fidelity becomes mode-centric for continuous variables, emphasizing regime-switching periodicity in addition to raw amplitude behavior.

5. Training objective, benchmarks, and empirical findings

The adversarial component uses a WGAN-style generator loss

$\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$6

The total generator objective is written conceptually as

$\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$7

with $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$8 implicit in the training schedule through the number of spectral update steps per epoch. Each epoch consists of $\text{Children}(w).\mathcal{A}=(u_1,\dots,u_T)$9 critic steps, one adversarial generator step, and $u_t\in\mathcal{V}(U)$0 generator steps on $u_t\in\mathcal{V}(U)$1 (Gueye et al., 30 Jun 2026).

To evaluate frequency-domain fidelity under exact ground truth, the paper introduces simulated relational datasets built from circulant categorical Markov chains. Two benchmark processes are used. The Noisy Cyclic Process (NCP) has

$u_t\in\mathcal{V}(U)$2

with $u_t\in\mathcal{V}(U)$3, producing sharp peaks at the fundamental frequency $u_t\in\mathcal{V}(U)$4 and harmonics as $u_t\in\mathcal{V}(U)$5. The Symmetric Sticky Process (SSP) has

$u_t\in\mathcal{V}(U)$6

and produces a low-pass spectrum that peaks at $u_t\in\mathcal{V}(U)$7 as $u_t\in\mathcal{V}(U)$8. In the relational simulation, the parent table contains 100 entities, each child sequence has length $u_t\in\mathcal{V}(U)$9, and the state space sizes are $w.\mathcal{A}$0 (Gueye et al., 30 Jun 2026).

The paper proposes two evaluation metrics. Spectral Density Divergence $w.\mathcal{A}$1 measures divergence between normalized mean spectral densities for continuous attributes, using KL in the reported experiments. Spectral Envelope Divergence $w.\mathcal{A}$2 measures the normalized $w.\mathcal{A}$3 distance between mean real and synthetic spectral envelopes for categorical attributes: $w.\mathcal{A}$4 These complement MSE(ACF), which measures mean squared error between real and synthetic autocorrelation functions (Gueye et al., 30 Jun 2026).

Empirically, Seq. RC-TGAN is reported to outperform SDV, ClavaDDPM, TimeGAN, and DoppelGANger on the spectral-envelope benchmarks. On simulated NCP data with $w.\mathcal{A}$5, $w.\mathcal{A}$6 is reported as $w.\mathcal{A}$7, compared with approximately $w.\mathcal{A}$8 for SDV and approximately $w.\mathcal{A}$9 for DoppelGANger. On simulated SSP with $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$0, Seq. RC-TGAN reports $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$1, compared with approximately $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$2 for the best baseline, TimeGAN. On Rossmann, the reported MSE(ACF) is $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$3 versus $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$4 for SDV; $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$5 is $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$6 versus $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$7 for DoppelGANger; and $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$8 is $\mathcal{A}_{cat}(U)=\{d_1,\dots,d_J\}$9 versus $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$0 for SDV. On Walmart, Seq. RC-TGAN reports MSE(ACF) $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$1, statistically tied with DoppelGANger’s $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$2, while improving $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$3 from $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$4 to $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$5 and $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$6 from $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$7 to $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$8 (Gueye et al., 30 Jun 2026).

The ablation study distinguishes recurrent modeling from spectral regularization. RC-TGAN and Seq. RC-TGAN without $\mathcal{A}_{cont}(U)=\{c_1,\dots,c_I\}$9 show similar $U$00 on simulated data, while adding the spectral envelope loss halves the envelope divergence in many cases, including NCP with $U$01, where the value drops from $U$02 to $U$03. On real data, recurrence alone improves MSE(ACF) and $U$04, whereas the full model is best across all metrics, particularly $U$05, which on Rossmann moves from $U$06 in RC-TGAN to $U$07 in the full model (Gueye et al., 30 Jun 2026).

6. Position within sequential GAN research and stated limitations

Seq. RC-TGAN belongs to a broader class of sequential GANs, but its design differs from several earlier paradigms. In “Learning to navigate image manifolds induced by generative adversarial networks for unsupervised video generation,” temporal generation is decoupled from static generation by first learning a frame generator and then training a recurrent video generator in the frozen latent manifold of that image model; recurrence therefore lives in latent space rather than directly in the final output generator (Albuquerque et al., 2019). In “ChainGAN: A sequential approach to GANs,” generation is organized as a base generator followed by a sequence of independently trained editor networks that iteratively refine a sample, again emphasizing staged generation and local objectives rather than a single monolithic adversarial mapping (Hossain et al., 2018).

By contrast, Seq. RC-TGAN uses an end-to-end conditional recurrent generator over relational child sequences, and its distinctive innovation is not merely sequentiality but the addition of a differentiable spectral envelope objective for categorical and VGM-discretized continuous time series. This places it at the intersection of relational synthesis, temporal GANs, and frequency-domain regularization. A plausible implication is that its main novelty lies less in recurrence itself than in identifying a frequency-domain criterion appropriate for mixed-type relational sequences.

The paper states several limitations. Computational cost is substantial because each batch and each feature require spectral density estimation and eigenvalue computations at multiple frequencies. The spectral envelope is defined for stationary processes and estimated on fixed-length windows, so strong non-stationarities or variable-length sequences may be harder to handle. The VGM assumption may be suboptimal for highly skewed or heavy-tailed continuous features, and mode assignment uses hard $U$08, so gradients do not pass through the discretization step. Large categorical cardinality $U$09 increases the dimension of $U$10 and the cost of generalized eigenvalue computations, which may raise runtime and numerical-stability concerns (Gueye et al., 30 Jun 2026).

The paper also outlines several extensions: replacing the RNN with transformers or temporal convolutions, matching fuller matrix spectra rather than only spectral envelopes, using alternative norms or divergences such as Wasserstein, replacing VGM with learned quantization or vector quantization, extending to irregular time grids, and handling richer relational schemas with deeper foreign-key structure or multiple child tables. These proposed directions suggest that Seq. RC-TGAN is best understood as a specific instantiation of a broader program: frequency-aware generative modeling for relational temporal data, with spectral envelopes providing the principal mechanism for preserving cyclic structure in categorical sequences (Gueye et al., 30 Jun 2026).