🤖 AI

L-GTA: Latent Generative Modeling for Time Series Augmentation

The paper introduces L-GTA, a Variational Autoencoder-based generative model that learns latent temporal representations and applies controlled perturbations with an equivariance objective to produce interpretable augmented data, significantly outperforming state-of-the-art methods in forecasting accuracy and transformation controllability across real-world time series datasets.

Original authors: Luis Roque, Vitor Cerqueira, Carlos Soares, Luis Torgo

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Luis Roque, Vitor Cerqueira, Carlos Soares, Luis Torgo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to predict the weather. You have a lot of data, but it's not enough to make the robot truly smart. You need more examples. But you can't just wait for more real weather to happen; that takes too long. So, you decide to create "fake" weather data to help the robot learn.

This is the problem of Data Augmentation. The paper introduces a new tool called L-GTA (Latent Generative Temporal Augmentation) to solve this problem for time-series data (data that changes over time, like stock prices, heart rates, or sales figures).

Here is how L-GTA works, explained through simple analogies:

1. The Problem with Old Methods

Traditionally, people made fake data by taking a real record and just "shaking" it up.

The Old Way: Imagine you have a recording of a song. To make a new version, you might just add static noise (jittering) or turn the volume up and down randomly (scaling).
The Flaw: This is like taking a photo and just blurring it. It's easy, but it doesn't capture the complex "soul" of the song. It's too simple and can create weird distortions that don't look like real life.

Newer methods use complex AI (like GANs or Diffusion models) to "dream up" new data.

The Flaw: These are like a chef trying to invent a new recipe by throwing random ingredients into a pot. Sometimes it tastes good, but you have no control over what it tastes like. You can't say, "Make it slightly saltier," and get a predictable result. It's also very slow and expensive to cook.

2. The L-GTA Solution: The "Master Blueprint"

L-GTA takes a different approach. Instead of shaking the final product or dreaming randomly, it works with the blueprint (the "latent space") of the data.

Think of a time series (like a stock chart) as a molded clay sculpture.

The Encoder (The Sculptor's Eye): L-GTA first looks at the real clay sculpture and figures out the internal structure—the "latent blueprint" that holds the shape together. It uses a special brain (Bi-LSTM with Attention) that understands both what happened in the past and what is happening right now.
The Latent Space (The Clay Itself): Instead of painting the outside of the sculpture, L-GTA works on the soft clay inside the mold.
The Transformation (The Sculptor's Hands): This is the magic part. The researchers can now gently squeeze, stretch, or twist the clay (the latent blueprint) in very specific, controlled ways.
- Want to add a little "jitter"? They wiggle the clay.
- Want to add a "drift" (a slow trend)? They slowly tilt the clay.
- Because they are working on the blueprint, the final sculpture (the new data) comes out looking natural and realistic, preserving the original style but with the new changes.

3. The "Equivariance" Rule: The Promise Keeper

The paper introduces a special rule called Equivariance. This is like a promise between the sculptor and the final statue.

The Rule: "If I twist the clay blueprint, the final statue must twist in the exact same way."
Why it matters: Without this rule, the AI might get confused. You might ask for a "stretch," and the AI might accidentally "shrink" the statue or make it wobble weirdly. L-GTA forces the AI to be consistent. If you apply a change to the blueprint, the output must reflect that change predictably. This makes the tool controllable and interpretable. You know exactly what you are getting.

4. The Results: Better, Faster, and Smarter

The authors tested L-GTA against other top methods (like TimeGAN and Diffusion-TS) using real-world data (tourism numbers, Wikipedia views, labor stats).

Prediction Power: When they used the fake data generated by L-GTA to train a prediction model, the model made 26% fewer mistakes than models trained with other fake-data methods. It even beat models trained on the original data alone by 27%.
Speed: L-GTA is incredibly fast. Other methods took hundreds or even thousands of times longer to generate the same amount of data.
Quality: The fake data looked just like the real data (high "fidelity") but also offered a wide variety of new scenarios (high "diversity").

Summary

L-GTA is like a smart, controllable clay sculptor. Instead of randomly guessing new shapes or just shaking the old ones, it understands the internal structure of the data. It allows you to gently tweak the "blueprint" to create new, realistic scenarios that are guaranteed to behave exactly as you intended. This helps computers learn faster and predict the future more accurately, without the high cost or lack of control found in other AI methods.

Technical Summary: L-GTA (Latent Generative Modeling for Time Series Augmentation)

Problem Statement

Time series analysis in domains such as finance, climate forecasting, and retail relies heavily on the quality and quantity of available data. However, real-world time series often suffer from limited labeled data, complex non-linear dynamics, noise, and irregular sampling. Traditional data augmentation techniques (e.g., jittering, scaling, warping) are often too simplistic to capture complex temporal dependencies and are difficult to apply in a controlled manner, potentially introducing artificial distortions that deviate from real-world scenarios.

While recent generative models (GANs, VAEs, Diffusion models) offer more realistic synthesis, they face significant challenges:

Training Instability: GANs often suffer from mode collapse and unstable optimization.
Computational Cost: Diffusion models require substantial resources due to iterative denoising.
Lack of Controllability: Most generative approaches rely on stochastic sampling, making it difficult to control specific characteristics of the generated samples or perform targeted augmentation.

Methodology: L-GTA

The authors propose L-GTA (Latent Generative Temporal Augmentation), a generative framework designed to produce controllable and interpretable time series augmentations. The model integrates a Variational Autoencoder (VAE) with a Bi-directional Long Short-Term Memory (Bi-LSTM) backbone and a temporal multi-head self-attention mechanism.

Architecture

Encoder: The input time series is processed by a Bi-LSTM to capture bidirectional temporal dependencies. A temporal multi-head self-attention layer is applied to these hidden states to capture long-range relationships. The encoder outputs Gaussian parameters ( $\mu_t, \Sigma_t$ ) for each timestep, representing a latent distribution.
Latent Space Manipulation: Instead of sampling new latent vectors from a prior distribution, L-GTA applies parametric transformations directly to the latent trajectory ( $v$ ). Supported transformations include jittering, scaling, magnitude warping, drift, and additive linear trends. These transformations can be chained sequentially.
Decoder: A Bi-LSTM decoder reconstructs the time series from the transformed latent trajectory ( $v'$ ). The decoder utilizes learned skip connections to propagate the transformations coherently, preserving the statistical structure of the original data.

Equivariance Regularization

A core innovation of L-GTA is the introduction of an equivariance objective. The model enforces consistency between transformations applied in the latent space and their effects in the observation (data) space. Formally, the model minimizes the difference between decoding a transformed latent vector and transforming a decoded series:
$D(T(v)) \approx T(D(v))$
This is achieved via an equivariance loss term ( $L_{equiv}$ ) added to the standard VAE reconstruction and KL divergence losses. This constraint ensures that the augmentation process is predictable and that the decoder maps latent perturbations to consistent changes in the reconstructed series.

Key Contributions

Controllable Generative Model: The proposal of L-GTA, a CVAE-based model that enables controlled augmentation by applying parametric transformations directly to latent trajectories, allowing for predictable and interpretable variations.
Equivariance Objective: The introduction of the first equivariance regularization objective specifically for controllable time series augmentation, enforcing consistency between latent and observation space transformations to improve reliability.
Comprehensive Evaluation: An extensive empirical study comparing L-GTA against direct transformation methods and State-of-the-Art (SOTA) generative models (TimeGAN, TimeVAE, Diffusion-TS) across multiple real-world datasets.

Experimental Results

The model was evaluated on five datasets (Tourism, Wiki2, Labour, M3) using downstream forecasting, distributional fidelity, and controllability metrics.

Downstream Forecasting: L-GTA consistently outperformed competing methods. In downstream forecasting tasks, it reduced the Mean Absolute Scaled Error (MASE) by approximately 26% compared to the strongest generative benchmark (TimeVAE) and 27% compared to using original data without augmentation. The improvement was even more significant (33–50%) compared to direct transformations and diffusion/GAN-based methods.
Controllability and Fidelity: L-GTA demonstrated high controllability, with latent transformation strength showing a strong monotonic correlation with the resulting distributional shift. Residual fingerprint analysis confirmed that the model preserves the structural signatures of intended augmentations (e.g., autocorrelation for smooth transformations, amplitude dependence for scaling).
Synthesis Quality: L-GTA achieved the best balance between fidelity and diversity, attaining near-perfect structural similarity (Cosine similarity $\approx$ 1.00) and the lowest distributional distances (Wasserstein, Fréchet, MMD) while maintaining high coverage of the real data manifold.
Computational Efficiency: L-GTA is significantly more efficient than SOTA generative baselines. It is approximately 8x faster than TimeVAE, 1,000x faster than Diffusion-TS, and 5,000x faster than TimeGAN.

Ablation Studies

Component analysis revealed that removing the equivariance regularization or simplifying the temporal encoder substantially degraded performance, leading to higher reconstruction errors (up to 166% relative to the baseline) and weaker controllability. Dynamic conditioning improved controllability but increased reconstruction error, suggesting its primary benefit lies in aligning with time-varying covariates rather than raw reconstruction fidelity.

Significance and Claims

The paper claims that L-GTA provides a substantially more effective approach to time series augmentation than existing generative or direct methods. By operating in the latent space with an equivariance constraint, L-GTA successfully bridges the gap between the realism of generative models and the controllability of traditional transformation techniques. The authors assert that the model produces synthetic data that is both realistic and diverse, significantly improving downstream predictive performance while remaining computationally efficient. The work is presented as the first to introduce equivariance regularization for this specific purpose, offering a robust solution for scenarios where labeled time series data is scarce.