Grasynda Achieves Improved Time Series Forecasting with Synthetic Data Augmentation

Scientists are increasingly focused on overcoming the challenges of limited data in time series forecasting, a critical issue when deploying deep learning models. Luis Amorim, Moises Santos, and Paulo J. Azevedo, from the Universities of Minho and Porto, alongside Carlos Soares and Vitor Cerqueira et al, present a new solution in their paper introducing Grasynda , a graph-based approach to synthetic time series generation. Unlike existing data augmentation techniques which often fail to maintain crucial data characteristics, Grasynda cleverly transforms time series into graph structures, encoding temporal dynamics within a transition probability matrix. Extensive evaluation on six benchmark datasets demonstrates that Grasynda consistently surpasses the performance of current state-of-the-art time series augmentation methods, offering a significant advancement for researchers and practitioners alike.

Scientists Background

Scientists have developed Grasynda, a novel graph-based method for generating synthetic time series data, addressing a critical need in Time series forecasting. Researchers recognised that deep learning architectures require substantial training datasets to generalise effectively, yet these are often unavailable in real-world applications, prompting the creation of this innovative technique. The core of Grasynda lies in its unique approach to transforming univariate time series into a network structure, representing each state as a node and transitions as directed edges, thereby capturing the series’ temporal dynamics within a transition probability matrix. This allows for the generation of new synthetic series by sampling transitions based on estimated probabilities, preserving the statistical properties of the original data and offering a significant advancement over existing methods.
The team achieved a breakthrough by encoding temporal dynamics into a transition probability matrix, effectively modelling the underlying data generation process. This graph representation inherently captures both local patterns, immediate state transitions, and global structures, overall network connectivity, providing a more holistic understanding of the time series data. Experiments demonstrate that Grasynda leverages this matrix to generate realistic synthetic data, significantly enhancing the performance of forecasting models, particularly when training data is limited. The method’s efficacy stems from its ability to maintain temporal dependencies and introduce realistic variability, overcoming the limitations of simpler augmentation techniques like jittering or scaling.

This study reveals that Grasynda consistently outperforms other time series data augmentation methods, including those employed in state-of-the-art time series foundation models such as Amazon’s Chronos. Researchers rigorously evaluated Grasynda using three neural network variations across six benchmark datasets, demonstrating its robustness and generalisability. The results indicate a clear improvement in forecasting accuracy when datasets are augmented with Grasynda-generated synthetic data, establishing its potential as a valuable tool for enhancing predictive performance. Furthermore, the method’s performance is competitive even when compared directly to other advanced data augmentation techniques.

The research establishes a new benchmark for synthetic time series generation, offering a sophisticated alternative to computationally intensive generative models like generative adversarial networks. By transforming time series into a graph representation, Grasynda avoids the pitfalls of mode collapse and overfitting often associated with these complex models, particularly when dealing with limited training data. The. This innovative approach encodes these dynamics within a transition probability matrix, enabling the generation of realistic synthetic data even when extensive datasets are unavailable.

Experiments demonstrate that Grasynda consistently outperforms other time series data augmentation methods, including those used in state-of-the-art time series foundation models. The team measured forecasting accuracy using six benchmark datasets and three distinct neural network architectures to validate Grasynda’s effectiveness. Results indicate a significant improvement in forecasting performance when datasets are augmented with Grasynda-generated synthetic time series. Specifically, the method leverages the transition probability matrix to sample transitions, maintaining the statistical properties of the original data and generating new series with preserved temporal dependencies.

This graph representation inherently captures both local patterns, immediate state transitions, and global structures, enhancing the quality of the synthetic data produced. Researchers discretized time series into finite states, representing transitions between these states as directed edges within a graph. The resulting transition probability matrix then served as the basis for generating new synthetic series by sampling transitions according to estimated probabilities. Tests prove that this process effectively captures the underlying data generation process, leading to more accurate forecasts. The study recorded consistent performance gains across all six benchmark datasets, showcasing the robustness and generalizability of the Grasynda approach.

Measurements confirm that Grasynda’s ability to preserve temporal dependencies and introduce realistic variability surpasses that of simpler methods like jittering or scaling. Furthermore, the breakthrough delivers competitive forecast accuracy compared to more computationally intensive generative models, such as generative adversarial networks. The method and all associated experiments are publicly available, facilitating further research and application of this promising technique. This work opens avenues for improved time series forecasting in scenarios where data scarcity is a significant challenge, potentially impacting fields ranging from finance to environmental monitoring.

Grasynda outperforms existing time series augmentation methods

Scientists have developed Grasynda, a new method for generating synthetic time series data to improve forecasting accuracy. The technique transforms univariate time series into a network structure, representing states as nodes and transitions as directed edges, then encodes temporal dynamics within a transition probability matrix. Extensive evaluations across six benchmark datasets, utilising three variations of the method, demonstrate that Grasynda consistently surpasses the performance of existing time series data augmentation techniques, even those integrated into state-of-the-art foundation models. This research establishes a novel approach to data augmentation that effectively preserves the underlying properties of time series data, addressing a key limitation of many current methods.

The consistent outperformance of Grasynda suggests it could be particularly valuable when limited training data is available, a common challenge in real-world forecasting applications. The authors acknowledge that the method’s complexity may require substantial computational resources for very large datasets. Future work could explore adapting Grasynda to multivariate time series or investigating its performance with different deep learning architectures, the code and experimental setup are publicly available to facilitate further research in this area.