Source-linked AI summary
Data Augmentation techniques in time series domain: A survey and taxonomy
Guillermo Iglesias, Edgar Talavera, Ángel González-Prieto, Alberto Mozo, Sandra Gómez-Canaval
TL;DR
Time-series deep-learning models depend on large, consistent datasets that are often scarce, incomplete, imbalanced, or privacy-constrained. The paper systematically reviews augmentation methods, evaluation metrics, and model problems, organizing representative approaches into a taxonomy. It concludes with a comprehensive overview of traditional methods, VAEs, and GANs, compared empirically across application areas and their advantages and disadvantages.
Problem
Time-series deep-learning training is constrained by datasets that are often incomplete, unbalanced, privacy-challenged, or difficult to obtain at sufficient scale.
Method
The paper systematically reviews time-series data-augmentation technologies, evaluation metrics, and model problems, organizing representative traditional, VAE, and GAN methods in a taxonomy.
Results
The review provides a comprehensive overview that compares representative augmentation methods empirically across application areas and highlights their advantages and disadvantages.
Takeaways & Limitations
The taxonomy and comparisons are intended to help future researchers position themselves and select time-series augmentation techniques.
Takeaways & Limitations
GAN evaluation lacks consensus on metrics, and many available metrics are centered on computer vision.
Abstract
from arXiv · showhide
With the latest advances in Deep Learning-based generative models, it has not taken long to take advantage of their remarkable performance in the area of time series. Deep neural networks used to work with time series heavily depend on the size and consistency of the datasets used in training. These features are not usually abundant in the real world, where they are usually limited and often have constraints that must be guaranteed. Therefore, an effective way to increase the amount of data is by using Data Augmentation techniques, either by adding noise or permutations and by generating new synthetic data. This work systematically reviews the current state-of-the-art in the area to provide an overview of all available algorithms and proposes a taxonomy of the most relevant research. The efficiency of the different variants will be evaluated as a central part of the process, as well as the different metrics to evaluate the performance and the main problems concerning each model will be analysed. The ultimate aim of this study is to provide a summary of the evolution and performance of areas that produce better results to guide future researchers in this field.
1 Introduction
Time-series deep-learning models face scarce, incomplete, unbalanced, and privacy-constrained datasets, while augmentation techniques require domain-specific analysis. The paper reviews augmentation technologies and their advantages and disadvantages for applying them to time-series data.
- Deep-learning models increasingly require larger datasets, making suitable training data more difficult to obtain.
- Time-series datasets are difficult to obtain because they are often limited by low availability, privacy issues, incompleteness, or imbalance.
- Data augmentation can increase dataset size through techniques such as preprocessing, noise, permutations, or synthetic-data generation.
- Time-series augmentation techniques require special attention because datasets differ substantially and each technique may suit different data.
- The paper reviews augmentation technologies and their positive and negative aspects to help researchers apply them to time-series data.
2 Problem statement
This survey addresses the need for a comprehensive view of techniques that generate or enhance time-series data. It aims to organize current approaches and clarify their differences and practical characteristics.
- The study reviews current techniques for generating new samples and enhancing temporal-series datasets.
- The survey seeks a complete view of how different approaches to time-series augmentation have developed.
- It focuses on traditional algorithms, VAEs, and GANs as three important pillars of current time-series data augmentation.
3 Related works
Prior reviews often focused on imaging, video, NLP, or specific time-series augmentation models, leaving no comprehensive comparison of all relevant techniques. This survey aims to close that gap with an updated, broad synthesis for future researchers.
- Existing augmentation reviews mainly target imaging, video, or NLP, whereas valid-data scarcity is especially pronounced in time series.
- Earlier time-series studies evaluated selected methods or focused on specific GAN taxonomies rather than the entire augmentation field.
- Some studies build synthetic traffic generators almost from scratch, requiring only an understanding of the data distribution.
- No previous study compiled all time-series augmentation technologies with a comprehensive comparison of their performance, results, advantages, and disadvantages.
- Because the field constantly evolves, updated reviews are needed to follow newly published techniques and current trends.
- The review brings together current time-series augmentation algorithms and contrasts their virtues, approaches, and differences.
4 Background
The section distinguishes direct transformation-based augmentation from generative approaches and introduces VAE and GAN mechanisms for synthesizing time-series data. It emphasizes that time-series-specific distributions constrain which augmentation techniques are valid and that excessive manipulation can distort signals.
- 4.1 Traditional algorithms: Time-series augmentation cannot directly reuse many computer-vision transformations because reorganizing portions can destroy the data tendency.Techniques must be adapted to time-series data or newly designed for particular datasets.
- 4.1 Traditional algorithms: Signal manipulation can distort data excessively, producing negative training outcomes.
- 4.1 Traditional algorithms: Traditional augmentation modifies input samples directly, whereas generative methods learn the data distribution to produce new samples.
- 4.2 Variational Autoencoder (VAE): VAEs encode inputs as probability distributions, sample latent points, and decode them to synthesize more diverse samples while reducing autoencoder overfitting.The latent distribution is represented by mean µ and standard deviation σ, which determine its center and output variability.
- 4.2 Variational Autoencoder (VAE): VAE training combines reconstruction error with a KL-divergence regularization term that organizes the latent distribution.
- 4.3 Generative Adversarial Networks (GANs): GANs use a Generator and Discriminator in competition: the Generator creates samples, while the Discriminator distinguishes real from synthetic data.Their interaction is formulated as a minimax game in which the Discriminator maximizes discrimination accuracy and the Generator minimizes it.
- 4.3 Generative Adversarial Networks (GANs): GAN training is difficult because the Nash equilibrium is rarely achieved under the architecture’s constant competition and instability.
5 Evaluation metrics
The survey reviews external performance, loss-based, distributional, and visual approaches for evaluating augmented time series. It emphasizes that no universal metric captures reliability, quality, and diversity across applications, so metrics should be adapted to the algorithm and field.
- No unique metric currently evaluates synthetic time-series reliability, quality, and diversity across all applications.
- GAN studies lack evaluation-metric consensus, while many available metrics were designed for computer vision.
- Metric selection should match the specific augmentation algorithm and application field.
- External performance evaluation compares models before and after augmentation, but measures model behavior rather than synthetic-sample quality directly.
- Forecasting studies use symmetric Mean Absolute Percentage Error and Mean Absolute Scaled Error to compare performance before and after augmentation.
- Classification studies compare accuracy across six neural architectures, but architecture-specific outcomes make the best augmentation algorithm difficult to identify and results hard to interpret.
- Distributional metrics directly compare real and synthetic probability distributions and can be applied regardless of the generation algorithm.
- The Wasserstein-Fourier distance measures time-series distribution differences through Wasserstein distance between frequency-domain energy distributions.
6 Data Augmentation algorithms review
The review organizes state-of-the-art time-series augmentation algorithms by their particularities, strengths, weaknesses, and relationships. It also cautions that applicability depends on the type of time series and target application.
- The section reviews state-of-the-art augmentation algorithms and groups related approaches into a proposed taxonomy.
- Algorithms differ in applicability: some target specific applications, whereas others are more general techniques.
6.1 Basic DA Methods
Basic augmentation methods transform existing time-series samples through operations such as slicing, noise addition, scaling, warping, and rotation. Their validity depends on preserving signal features and application-specific structure.
- Traditional methods create synthetic samples by directly deforming, shortening, enlarging, or otherwise modifying existing data.
- Slicing: Slicing crops a selected window from a time series, but may remove important features and produce invalid synthetic samples.
- Concatenating and resampling: Concatenating and resampling detects characteristic points before recombination to preserve signal features, but requires a domain-specific detector.
- Jittering: Jittering adds typically Gaussian noise at each time step, with noise parameters controlling the deformation magnitude and shape.
- Jittering: Jittering must be adapted to each case because it produced negative learning in at least one wearable-sensor application.
- Scaling: Scaling changes signal magnitude while preserving overall shape; homogeneous scaling multiplies values by α, where α > 0 controls change magnitude.
- Warping: Magnitude warping applies smooth, pointwise scaling through cubic-spline interpolation, but assumes transformed samples remain valid.
- Warping: Time warping changes temporal structure by stretching or shortening time slices rather than altering signal magnitude.
6.2 Data Augmentation through VAE
VAE-based augmentation learns structured latent representations to generate or reconstruct time-series data for tasks including anomaly detection and imputation. Architectures and hyperparameters specialize to dataset formats and downstream objectives.
- VAEs extend autoencoder-based generation by controlling deviation from original data and producing synthetic samples for time-series applications.
- VAE configurations specialize according to dataset format and downstream use, including classification, forecasting, imputation, and prediction.
- The reviewed VAE architectures are grouped schematically by research direction and application.
- Anomaly detection: In anomaly detection, VAEs generate additional anomalous samples when available invalid examples are insufficient for model training.
- Anomaly detection: A conditional VAE with LSTM layers learns normal versus anomalous ECG samples for classification.
- Anomaly detection: Smoothness-Inducing Sequential VAE uses recurrent layers and corrective time-step biases to reduce abrupt changes in synthesized sequences.
- Data imputation: For imputation, VAEs generate information on demand to fill temporal gaps according to the original data distribution.
- Data imputation: GlowImp combines a Glow VAE and Wasserstein GAN: latent interpolation handles missing values, while adversarial discrimination promotes realistic samples.
6.3 Data Augmentation through GAN
GAN-based augmentation learns or synthesizes time-series data, either by adapting architectures to temporal structure or by transforming sequences into images. TimeGAN and Conditional Sig-Wasserstein GANs specifically address temporal dependencies through specialized objectives or metrics.
- GANs learn data distributions by extracting sample features rather than directly copying them, enabling synthetic data generation for augmentation.
- Recurrent GANs: Recurrent GANs use recurrent layers to retain temporal links, with LSTM-based variants applied to music synthesis, ECG generation, and anomaly detection.
- Image-based GANs: Image-based approaches transform time-series or signal data into images, allowing conventional image GANs to generate synthetic samples before reversing the transformation.
- TimeGAN: TimeGAN adds a supervised loss that compares generator forecasts with real data to learn stepwise dependencies in sequences.
- Conditional Sig-Wasserstein GAN: Conditional Sig-Wasserstein GAN uses a path-space Wasserstein metric and replaces the discriminator neural network with linear regression to maintain long temporal dependencies.
6.4 DA based on Dynamic Time Warping (DTW)
DTW-based augmentation manipulates time-series manifolds or warping paths to create new sequences. The reviewed methods differ in how they select neighboring samples, weight them, and introduce controlled path or signal variation.
- DTW measures similarity between sequences and supports augmentation evaluated with augmented training sets, 1-NN classification, and 85 UCR datasets.
- Weighting methods: Average All uses flat Dirichlet weights but can fill data spaces where valid samples should not occur.
- Weighting methods: Average Selected restricts interpolation to close samples on the manifold, reducing the risk of populating empty regions.
- SPAWNER: SPAWNER modifies DTW warping paths by inserting a random point, constraining alignment with ξ, and concatenating paths to generate new signals.
- SPAWNER: For some multivariate signals, SPAWNER additionally applies normally distributed random variance to each generated signal point.
7 Discussion
The discussion contrasts traditional, VAE, and GAN-based augmentation by control, diversity, flexibility, and training difficulty. It also identifies validation and reliability as persistent challenges for generative models.
- Traditional algorithms: Traditional algorithms modify existing examples with few hyperparameters, reducing setup time and training-data requirements while providing controlled variation.Their main limitation is that they cannot generate genuinely new elements and may produce invalid or lower-quality examples.
- VAE algorithms: VAEs provide the greatest control over generated-data variability by directly influencing the latent distribution’s standard deviation.They remain useful for anomaly detection and applications requiring precise variability control, despite generating less data than newer generative networks.
- GAN algorithms: GANs learn the original dataset’s distribution to generate synthetic data with better generalisation, but sacrifice some control over generation.Their recent research attention has supported continued efforts to improve their results.
- Open issues and challenges: GAN training is especially complex because mode collapse, instability, convergence difficulties, gradient vanishing, and Nash-equilibrium issues complicate optimisation and reliability.The absence of a clear stopping condition further makes it difficult to determine when generated data are high quality.
- Generative-model flexibility: Neural-network generative models can add dataset diversity and reuse architectures across disparate problems, unlike traditional methods that often require algorithm redesign.Examples include one architecture applied to different datasets and LSTM-GAN used for electrocardiograms and taxi statistics.
- Open issues and challenges: Generative-model validation remains unsettled because the research community has no consensus on which evaluation metrics or stopping criteria to use.Studies employ diverse combinations of empirical, visualisation, discriminative, predictive, and error-based measures.
8 Conclusion
The paper presents a comprehensive taxonomy of time-series data-augmentation algorithms and empirically compares representative methods across application areas. It uses the comparison to identify advantages, disadvantages, and directions for future research.
- Conclusion: The review organises time-series data augmentation into basic and advanced approaches covering traditional algorithms, VAEs, and GANs.It summarises representative methods within each algorithmic family.
- Conclusion: Representative augmentation methods are compared empirically and disaggregated by application area.The review also highlights each approach’s advantages and disadvantages.
- Conclusion: The taxonomy and comparative synthesis are intended to support future research on data augmentation for time series.The paper frames this need within the field’s significant recent evolution.