Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Are Synthetic Time-series Data Really not as Good as Real Data?

Fanzhe Fu, Junru Chen, Jing Zhang, Carl Yang, Lvbin Ma, Yang Yang

TL;DR

The paper tackles the limited generalization of time-series models by introducing InfoBoost, a universal, non-deep-learning data synthesis framework that generates synthetic time-series from explicit components: multi-source rhythmic data (MRD), noise (TN&NR), and trend information (TI). It also trains a universal representation extractor solely on synthetic data to decompose MRD, TN&NR, and TI, enabling explicit, interpretable features without real data or fine-tuning. Across 35 real datasets, models trained on InfoBoost synthetic data consistently outperform those trained on real data in reconstruction tasks, demonstrating strong cross-domain generalization. This approach opens opportunities for data-efficient, unsupervised or self-supervised time-series analysis with broad practical impact in finance, energy, weather, healthcare, and beyond.

Abstract

Time-series data presents limitations stemming from data quality issues, bias and vulnerabilities, and generalization problem. Integrating universal data synthesis methods holds promise in improving generalization. However, current methods cannot guarantee that the generator's output covers all unseen real data. In this paper, we introduce InfoBoost -- a highly versatile cross-domain data synthesizing framework with time series representation learning capability. We have developed a method based on synthetic data that enables model training without the need for real data, surpassing the performance of models trained with real data. Additionally, we have trained a universal feature extractor based on our synthetic data that is applicable to all time-series data. Our approach overcomes interference from multiple sources rhythmic signal, noise interference, and long-period features that exceed sampling window capabilities. Through experiments, our non-deep-learning synthetic data enables models to achieve superior reconstruction performance and universal explicit representation extraction without the need for real data.

Are Synthetic Time-series Data Really not as Good as Real Data?

TL;DR

The paper tackles the limited generalization of time-series models by introducing InfoBoost, a universal, non-deep-learning data synthesis framework that generates synthetic time-series from explicit components: multi-source rhythmic data (MRD), noise (TN&NR), and trend information (TI). It also trains a universal representation extractor solely on synthetic data to decompose MRD, TN&NR, and TI, enabling explicit, interpretable features without real data or fine-tuning. Across 35 real datasets, models trained on InfoBoost synthetic data consistently outperform those trained on real data in reconstruction tasks, demonstrating strong cross-domain generalization. This approach opens opportunities for data-efficient, unsupervised or self-supervised time-series analysis with broad practical impact in finance, energy, weather, healthcare, and beyond.

Abstract

Time-series data presents limitations stemming from data quality issues, bias and vulnerabilities, and generalization problem. Integrating universal data synthesis methods holds promise in improving generalization. However, current methods cannot guarantee that the generator's output covers all unseen real data. In this paper, we introduce InfoBoost -- a highly versatile cross-domain data synthesizing framework with time series representation learning capability. We have developed a method based on synthetic data that enables model training without the need for real data, surpassing the performance of models trained with real data. Additionally, we have trained a universal feature extractor based on our synthetic data that is applicable to all time-series data. Our approach overcomes interference from multiple sources rhythmic signal, noise interference, and long-period features that exceed sampling window capabilities. Through experiments, our non-deep-learning synthetic data enables models to achieve superior reconstruction performance and universal explicit representation extraction without the need for real data.
Paper Structure (17 sections, 3 equations, 8 figures, 1 table)

This paper contains 17 sections, 3 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related works
  3. Methodology
  4. Universal Synthetic time-series data
  5. Generating Multi-source Rhythmic Data
  6. Generating Different Types of Noise
  7. Generating Trend Information
  8. Signal-to-Noise-and-Trend Ratio
  9. Universal Time Series Representation Extraction
  10. Synthetic info normalization
  11. Representation Extraction Training
  12. Experiments
  13. Reconstruction of Real Data Solely Trained on Synthetic Data
  14. A Fair Fight with Real Data
  15. Giving It All with Synthetic Data
  16. ...and 2 more sections

Figures (8)

  • Figure 1: Schematic of the InfoBoost model illustrating the data synthesis, representation learning, and representation prediction processes. In the diagram, the label 'Rhyth' corresponds to the multi-source rhythmic data (MRD), 'Noise' corresponds to different types of noise and their noise ratios (TN & NR), and 'Trend' corresponds to trend information (TI). This visual representation elucidates the individual roles of each component within the InfoBoost framework. 'Sync' stands for 'synthesized data', representing data that is artificially generated, integrating MRD, TN & NR, and TI.
  • Figure 2: This image illustrates a possible set of five corresponding sine waves, each obtained by random sampling of frequency phases and amplitudes within their respective ranges, and it should be noted that the number of sine waves is also randomly determined. Additionally, the image showcases the composite rhythmic data generated by the superposition of these randomly determined sine waves.
  • Figure 3: The list of noise distributions along with their corresponding categories. The "Category" column specifies the category each noise type belongs to. The categories are abbreviated as follows: CCD (Common Continuous Distributions), CDD (Common Discrete Distributions), HTD (Heavy-Tailed Distributions), DRND (Distributions Related to Normal Distribution), and SPD (Shape Parameter Distributions).
  • Figure 4: Normalization of parameters (norm params) used in the data synthesis process. The normalized parameters are organized into a multi-channel matrix, aligning with the sampling window length of the synthetic data.
  • Figure 5: The experimental results in \ref{['sameNumRecon']} show that smaller values for all metrics indicate better performance. Here, SDL represents structural dissimilarity loss, DTW represents dynamic time warping distance, and DH represents distance between histograms, MSE represents Mean Squared Error. To demonstrate the reconstruction performance directly across various domains of time-series data, we categorized all tested time-series data into five major groups. We computed the average losses based on each dataset and then calculated the average losses for all datasets within each major group, this allows us to observe relatively reasonable model performance across vastly different types of datasets, even when there is a significant difference in data volume.
  • ...and 3 more figures