TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective

AI summary

Overview

Research area: Generative machine learning for multivariate time series — specifically diffusion models, graph neural networks, and synthetic data evaluation. Technical level: Advanced (assumes familiarity with diffusion probabilistic models, graph convolutional networks, and variational autoencoders). Scope: The paper proposes TSGDiff, a framework that converts time series into dynamic graphs, diffuses over their latent graph representations, and introduces a new graph-aware evaluation metric called Topo-FID.

What This Paper Is About

Generating realistic synthetic multivariate time series is hard because models must reproduce both temporal dependencies (how values evolve over time) and the structural relationships between variables. Most existing methods, including GANs, VAEs, and diffusion models like Diffusion-TS, operate directly on raw numeric sequences and inherit a Euclidean-space assumption that the authors argue limits their ability to represent topological structure. TSGDiff instead treats a time series window as a graph — built from Fourier-spectrum periodicity and consecutive-step links — and runs the diffusion process inside a learned graph latent space.

Key Contributions

1. TSGDiff framework. The authors describe it as the first framework to rethink synthetic time series generation from a graph-based perspective, integrating a diffusion model within a latent graph space to model the structural representation distribution of time series.
2. Topological Structure Fidelity (Topo-FID). A graph-aware evaluation metric that combines Graph Edit Similarity (adjacency-matrix differences) and Structural Entropy Similarity (entropy of node degree distributions), with a default weighting of α = 0.5 between the two sub-metrics.
3. Fourier-based dynamic graph construction. Edges are created between consecutive time steps and between nodes identified as periodic neighbors via the top three highest-amplitude frequencies detected in each window's spectrum, intended to capture short-, medium-, and long-term variation patterns.
4. Empirical validation. Experiments on six real-world datasets comparing against Diffusion-TS, TimeGAN, Cot-GAN, and TimeVAE, plus an ablation study on ETTh and an efficiency analysis.

Main Findings

• Topo-FID favors TSGDiff on all six datasets. TSGDiff scores 0.986 (ETTh), 0.826 (Stocks), 0.869 (Exchange), 0.867 (Weather), 0.869 (Wind), and 0.852 (EEG). The next-best scores are Diffusion-TS at 0.864 (ETTh), 0.785 (Stocks), 0.806 (Exchange), 0.828 (Weather), 0.819 (Wind), and 0.788 (EEG); Cot-GAN reaches 0.894 on ETTh, and TimeVAE reaches 0.904 on ETTh.
• Discriminative score improvement. The paper states TSGDiff improves the average Discriminative score by 50% across the six datasets. On ETTh, TSGDiff records 0.056 versus 0.111 for Diffusion-TS, 0.353 for TimeGAN, 0.114 for Cot-GAN, and 0.346 for TimeVAE.
• Context-FID and Correlational results. TSGDiff posts Context-FID values of 0.224 (ETTh), 0.357 (Stocks), 0.061 (Exchange), 0.353 (Weather), 0.256 (Wind), and 0.020 (EEG). Correlational scores are 0.024, 0.026, 0.019, 0.035, 0.022, and 0.201 respectively.
• Predictive scores are low across the board. TSGDiff achieves 0.020 (ETTh), 0.005 (Stocks), 0.004 (Exchange), 0.009 (Weather), 0.007 (Wind), and 0.001 (EEG), with Diffusion-TS close behind at 0.026, 0.005, 0.005, 0.011, 0.009, and 0.002.
• Ablation: removing KL divergence is most damaging. On ETTh, dropping the KL loss collapses Topo-FID from 0.986 to 0.787 and sends Context-FID from 0.224 to 44.467. Removing the denoising loss gives Topo-FID 0.908 and Context-FID 3.922; removing the Fourier loss gives Topo-FID 0.883 and Context-FID 0.407.
• Graph construction is computationally cheap. Graph construction takes 3.12 s (ETTh), 0.82 s (Stocks), and 3.79 s (Weather), representing 0.07%, 0.08%, and 0.09% of total training time of 4,506.15 s, 970.23 s, and 4,025.31 s respectively, across 68,000, 14,500, and 61,000 iterations. The paper reports costs consistently under 5 seconds and under 0.1% of training time.
• Visualization. Using t-SNE projections and kernel density estimation on ETTh and Wind, the paper reports that the first row (t-SNE) does not clearly indicate which method performs better, while the second row (kernel density estimation) shows TSGDiff's synthetic distribution closely matches the original data, outperforming Diffusion-TS.

Methodology in Plain English

The pipeline has three stages. First, long time series are cut into fixed-length windows (window size 48, stride 1) and Min-Max normalized to [0, 1]. Each window is transformed with a discrete Fourier transform per variable; the top three frequency peaks are converted into periods, and those periods determine which time steps get connected as periodic neighbors. Every consecutive pair of steps is also always connected, producing an adjacency matrix.

Second, a graph convolutional encoder with three stacked GraphConvBlocks (hidden and latent dimensions of 1600) aggregates neighbor features, applies mean pooling, and outputs a mean and log-standard-deviation. A latent vector is sampled with the reparameterization trick, giving a VAE-style latent space regularized by a KL divergence loss toward a standard normal.

Third, a diffusion process operates on that latent vector rather than on the raw data. The forward process adds Gaussian noise over 1000 timesteps; the reverse process is a denoising network of three DiffusionBlocks with 64 units each (each block a three-layer MLP). A graph decoder of four fully connected layers with Tanh output restores the 1600-dimensional features back to the original dimension.

Training minimizes a weighted sum: reconstruction loss plus β·KL (β = 0.2) plus γ·denoising (γ = 1) plus δ·Fourier loss (δ = 1), where the Fourier term, inspired by Diffusion-TS, aligns the frequency content of reconstructions with the original data. Settings: batch size 128, learning rate 0.01, 500 epochs, five repeated runs on an NVIDIA RTX 4090 GPU with 24GB memory, implemented in PyTorch.

For evaluation, Topo-FID averages α·Graph Edit Similarity + (1−α)·Structural Entropy Similarity, with α = 0.5, Graph Edit Similarity defined as 1 minus the normalized element-wise absolute difference between real and synthetic adjacency matrices over N², and Structural Entropy Similarity defined as 1/(1 + |H(A) − H(Â)|) using degree-distribution entropy with ε = 10⁻¹⁰.

One point of ambiguity in the text: the abstract and introduction describe graph nodes as representing variables, whereas Section 3.2 states that each time step in a slice corresponds to a node with the time series value as its feature. The paper does not reconcile these two descriptions.

Why This Matters

Impact on research. The paper argues that existing generative models fail to capture structural and topological properties because they process raw sequences directly, and that Topo-FID fills a gap in evaluation since Context-FID, Correlational, Discriminative, and Predictive scores do not measure structural or frequency characteristics. It positions graph-based generative modeling as an under-explored direction for time series synthesis, in contrast to graph methods that have mainly been used for spatiotemporal prediction.

Real-world applications (domains named in the paper as users of multivariate time series generation):

• Energy management
• Financial market forecasting
• Medical monitoring
• Wind farm generation data analysis (one of the benchmark domains)

Industry relevance. Synthetic time series can substitute for scarce, private, or sensitive real data in downstream model training. TSGDiff's reported graph construction overhead — under 5 seconds and under 0.1% of training time — suggests the graph conversion step adds little to an existing diffusion training budget, which matters for practical deployment at scale.

Future Directions

• The paper concludes by framing TSGDiff as "a foundation for future research into scalable, adaptive, and high-fidelity generative models," but does not name specific follow-up research questions.
• Extending the graph construction beyond the current fixed rules (always linking consecutive steps, plus the top three frequency peaks detected by a peak-detection method) is not explored; the paper does not report experiments with alternative edge-construction strategies.
• The sensitivity of Topo-FID to its α weighting is not examined — α is fixed at 0.5 by default with no reported variation study.
• The sliding window size is fixed at 48 across all datasets with a stride of 1, and the paper does not report whether performance changes with other window sizes or strides.

Target Audience

Researchers and practitioners in generative modeling, time series analysis, and graph machine learning who are interested in non-sequence representations of temporal data or in evaluation metrics for synthetic structured data. The paper is also relevant to applied scientists working with energy, finance, weather, wind, and EEG data who need to generate privacy-preserving or augmentation data with faithful structural properties. Readers should already understand diffusion probabilistic models and graph neural networks, as both are assumed rather than introduced in depth.

Authors’ abstract

Diffusion models have shown great promise in data generation, yet generating time series data remains challenging due to the need to capture complex temporal dependencies and structural patterns. In this paper, we present \textit{TSGDiff}, a novel framework that rethinks time series generation from a graph-based perspective. Specifically, we represent time series as dynamic graphs, where edges are constructed based on Fourier spectrum characteristics and temporal dependencies. A graph neural network-based encoder-decoder architecture is employed to construct a latent space, enabling the diffusion process to model the structural representation distribution of time series effectively. Furthermore, we propose the Topological Structure Fidelity (Topo-FID) score, a graph-aware metric for assessing the structural similarity of time series graph representations. Topo-FID integrates two sub-metrics: Graph Edit Similarity, which quantifies differences in adjacency matrices, and Structural Entropy Similarity, which evaluates the entropy of node degree distributions. This comprehensive metric provides a more accurate assessment of structural fidelity in generated time series. Experiments on real-world datasets demonstrate that \textit{TSGDiff} generates high-quality synthetic time series data generation, faithfully preserving temporal dependencies and structural integrity, thereby advancing the field of synthetic time series generation.