DoppelGANger: Time-Series & Metadata Synthesis

• DoppelGANger is a generative adversarial network that uses dual-stream generators to synthesize multi-dimensional time-series and categorical metadata.
• It leverages batch RNNs and per-sample auto-normalization alongside dual discriminators to capture temporal dependencies and preserve metadata correlations.
• The framework is designed for high-fidelity, privacy-sensitive synthetic data generation for applications like financial modeling and econometric forecasting.

DoppelGANger (DGAN) is a generative adversarial network (GAN) framework explicitly designed for flexible, high-fidelity synthesis of multi-dimensional time-series data in conjunction with static or categorical metadata. DGAN integrates dual-stream generation for time-series and metadata, batch RNN construction, per-sample auto-normalization, and auxiliary discrimination to target scenarios ranging from financial transaction simulation to networked-systems trace sharing and econometric data synthesis (Karst et al., 2024, Dannels, 2023, Lin et al., 2019). The architecture and training methodology are motivated by the limitations of classical GANs in capturing temporal dependencies, joint measurement–metadata correlations, and rare-event synthesis, as well as the need for privacy-preserving data sharing where classic anonymization falls short.

1. Core Architecture and Conditioning Paradigm

DGAN's core innovation is its modular separation of generation streams: (i) a generator G_t for time-series feature blocks and (ii) a generator G_m for static metadata or categorical attributes. Both streams are conditioned on a shared latent variable z (sampled from p_z) and relevant covariates c (typically embedded and concatenated), ensuring the output sequences exhibit the correct cross-dependencies and temporal/categorical structure. For dataset classes with explicit multivariate measurements and metadata—network traces, transaction logs, econometric segments—the time-series generator is implemented as a batched LSTM, while the metadata generator is typically a multilayer perceptron (MLP) (Lin et al., 2019, Dannels, 2023).

Batch generation is fundamental: DGAN produces entire sequences (length T) or blocks of consecutive time points per RNN pass, reducing the effective unroll depth, enabling robust learning of long-term dependencies (e.g., autocorrelations, regime switches). Auto-normalization is applied per-sample, with learnable [min, max] normalization vectors generated alongside metadata: the time series for each synthetic "trace" is thus produced as a normalized vector and then scaled into realistic dynamic ranges by the auxiliary generator outputting \widehat{min} and \widehat{max} (Lin et al., 2019, Dannels, 2023).

Dual discriminators are used:

• The primary discriminator D or D_1 distinguishes real vs. generated pairs of (time-series, metadata).
• The auxiliary discriminator D_aux or D_2 operates solely on the metadata slice, preventing meta–mode collapse and ensuring correct marginal distributions (Lin et al., 2019, Dannels, 2023).

2. Training Objective, Loss Functions, and Optimization

DGAN is trained in an adversarial minimax framework, using dual losses for both the generators and discriminators. The loss functions capture both standard GAN optimization and auxiliary terms specific to metadata reconstruction: $$\mathcal{L}_{D_m} = -\,\mathbb{E}_{(m,c)\sim p_{\mathrm{data}}}[\log D_m(m\mid c)] -\,\mathbb{E}_{z\sim p_z,\;c\sim p_c}[\log(1 - D_m(G_m(z,c)\mid c))]$$

$$\mathcal{L}_{D_t} = -\,\mathbb{E}_{(x_{1:T},c)\sim p_{\mathrm{data}}}[\log D_t(x_{1:T}\mid c)] -\,\mathbb{E}_{z\sim p_z,\;c\sim p_c}[\log(1 - D_t(G_t(z,c)\mid c))]$$

$$\mathcal{L}_{G_m} = -\,\mathbb{E}_{z,c}[\log D_m(G_m(z,c)\mid c)] + \lambda_{\mathrm{rec}}\mathbb{E}_{(m,c)\sim p_{\mathrm{data}},\:z}[\|G_m(z,c)-m\|^2_2]$$

$$\mathcal{L}_{G_t} = -\,\mathbb{E}_{z,c}[\log D_t(G_t(z,c)\mid c)]$$

Total generator loss: $$\mathcal{L}_G = \mathcal{L}_{G_m} + \mathcal{L}_{G_t}$$ In other instantiations (notably (Lin et al., 2019, Dannels, 2023)), the losses are Wasserstein-1 with gradient penalty and a weighted auxiliary loss steering the metadata distribution.

Alternating updates are performed (typically 1:1 or 5:1 discriminator:generator) using Adam optimizer (default learning rate 1e-4–2e-4, \beta_1=0.5), with model-specific early-stopping or fixed epoch schedules (e.g., 50–100 epochs for financial data (Karst et al., 2024), up to 2000 epochs in economic time-series synthesis (Dannels, 2023)). Auto-normalization, dropout regularization in discriminators, and "generation flags" for variable-length trace output are essential stabilizers (Lin et al., 2019).

3. Quantitative Performance: Fidelity, Privacy, Efficiency

Financial Transaction Synthesis

On large-scale transaction data (5.2M rows real/4.2M IBM-simulated), DGAN sources full-sequence batches (B traces of length T) for training. Empirical evaluation uses structural, statistical, and privacy metrics:

Algorithm	Col-wise Fid. (KS)	Row-wise Fid. (Pearson)	Synthesis Novelty	DCR (5%)	NNDR (5%)	Time (s)	NetSimile
CTGAN	0.8779	0.9526	0.99998	0.0197	0.9855	2267	30.96
DGAN	0.4781	0.8620	1.00000	1.8112	0.9412	743	30.78
WGAN	0.2406	0.9451	1.00000	1.7756	0.9999	538	NaN
FinDiff	0.9543	0.9852	0.84101	0.0003	0.9878	625	30.94
TVAE	0.8968	0.9737	0.99922	0.00997	0.9799	401	31.20

Privacy is assessed via: $$\mathrm{DCR}_{5\%} = \mathrm{quantile}_{0.05}\Bigl(\bigl\{\min_{y_j}\|x_i - y_j\|_2\Bigr\}_{i=1}^N\Bigr)$$ DGAN achieves the largest DCR (1.8112) and perfect synthesis novelty (1.0), supporting its role in privacy-sensitive environments. However, it lags in column-wise (KS) and row-wise (Pearson) fidelity compared to FinDiff and TVAE. Efficiency is mid-tier: substantially faster than CTGAN (743s vs. 2267s) but slower than TVAE or WGAN.

Time-Series and Predictive Workloads

In forecasting and classification settings (e.g., Treasury yields, recession flags), DGAN-generated synthetic data boosts LSTM performance:

• 1-day-ahead forecasts: Synthetic-trained LSTM RMSE (0.046 for 1-year, 0.062 for 10-year) outperforms real-only in short-horizon regimes.
• Combined training (real + synthetic): Lowest MAPE for 10-year Treasury (3.48%), best longer-horizon results.
• Recession classification: LSTM-AUC improves from 0.69 (real) to 0.81 (combined) and 0.86 (synthetic). Logistic regression is unchanged, confirming LSTM models benefit from synthetic augmentation (Dannels, 2023).

4. Comparison with Related GAN-Based and Statistical Methods

DGAN was devised to address fidelity failures endemic to vanilla GANs, naive MLPs, and teacher-forced RNNs—specifically, failures in reproducing autocorrelation, scale diversity, metadata histograms, and joint measurement–metadata CDFs. Quantitative benchmarks on public systems and economic datasets show:

• DGAN achieves MSE=0.0009 on autocorrelation (WWT), outperforming RCGAN (0.0103) and AR (0.2777).
• Cross-domain correlation (GCUT): Synthetic Pearson statistics closely align with real data.
• Predictive modeling (GCUT classification): Up to 43% higher accuracy than AR baseline.
• Spearman algorithm ranking: 1.00 (perfect preservation) on GCUT, higher than all GAN and non-GAN baselines.
• Training time: DGAN is substantially more efficient than comparably expressive methods, requiring ~17h on 50k samples (Tesla V100 GPU), versus 258h (TimeGAN) or 29h (RCGAN) (Lin et al., 2019).

In financial data benchmarks (Karst et al., 2024), DGAN's unique privacy advantage distinguishes it from FinDiff (diffusion), TVAE (VAE), CTGAN, and vanilla WGAN. CTGAN achieves the best overall trade-off between fidelity and privacy; DGAN is the recommended choice where privacy risk is paramount.

5. Limitations and Open Problems

Privacy remains unresolved in a formal sense: membership inference attacks and overfitting risks persist except at large data scales. Differentially private GAN training (gradient clipping + noise) causes catastrophic degradation of temporal structure and autocorrelation—rendering DP-GAN approaches impractical for time-series (Lin et al., 2019). DGAN assumes the measured trace–metadata joint distribution P(R|A) is stationary and does not extrapolate to unseen metadata regimes or simulate stateful causal protocols directly.

In heterogeneously scaled or long-memory data (very large T), fixed windowing strategies can undercapture long-range dependence, and model performance is sensitive to window size and normalization/auto-scaling strategies (Dannels, 2023). Surplus synthetic data can induce noise and spurious decision boundaries in downstream tasks, especially under limited real-sample regimes.

6. Practical Usage and Recommendations

In financial applications, DGAN excels in inter-bank data sharing or any setting where re-identification risk is the dominant concern (Karst et al., 2024). It outperforms WGAN in graph-structure preservation (NetSimile), though none of the tested methods fully reconstruct complex transaction graphs. For tasks prioritizing high-fidelity data replication or upsampling (e.g., data augmentation, distribution matching), FinDiff and TVAE are superior. CTGAN provides the overall most balanced performance across privacy, fidelity, and synthesis novelty.

Key factors for practical deployment:

• Use per-sample auto-normalization and dual discriminators to avoid mode collapse and meta-distribution artifacts.
• Exploit batch generation at block sizes S such that T/S ≈ 50 for stable LSTM training.
• Monitor privacy metrics (DCR and NNDR) and novelty—DGAN attains optimal separation from real data at the cost of some realism in marginal distributions.
• Tune batch size, sequence length, and learning rates to fit the data regime; grid search is recommended as full prescriptions are dataset-specific (Karst et al., 2024).

A plausible implication is that, for large-scale multi-dimensional time-series with structured metadata, DGAN (or its variants) provides a robust, modular workflow for synthetic data generation, with demonstrated utility in financial, networked systems, and econometric forecasting contexts (Karst et al., 2024, Dannels, 2023, Lin et al., 2019).