AutoDA-Timeseries: Automated Data Augmentation for Time Series

Zijun Dou1, Zhenhe Yao1, Zhe Xie1, Xidao Wen2, Tong Xiao1, Dan Pei1
1Tsinghua University
2Aliaba Cloud Computing Company
Accepted at ICLR 2026
Paper Code

Abstract

Data augmentation is a fundamental technique in deep learning, widely applied in both representation learning and automated data augmentation (AutoDA). In representation learning, augmentations are used to construct contrastive views for learning task-agnostic embeddings, while in AutoDA the augmentations are directly optimized to improve downstream task performance. However, existing paradigms face critical limitations: representation learning relies on a two-stage scheme with limited adaptability, and current AutoDA frameworks are largely designed for image data, rendering them ineffective for capturing time series–specific features. To address these issues, we introduce AutoDA-Timeseries, the first general-purpose automated data augmentation framework tailored for time series. AutoDA-Timeseries incorporates time series features into augmentation policy design and adaptively optimizes both augmentation probability and intensity in a single-stage, end-to-end manner. We conduct extensive experiments on five mainstream tasks, including classification, long-term forecasting, short-term forecasting, regression, and anomaly detection, showing that AutoDA-Timeseries consistently outperforms strong baselines across diverse models and datasets.

Representation Learning vs. AutoDA

MY ALT TEXT

Existing applications of data augmentation in time series analysis can be broadly categorized into two paradigms: representation learning and automated data augmentation (AutoDA). In the representation learning paradigm, augmentations are primarily used to construct positive and negative views for contrastive learning. An encoder is first optimized to learn task-agnostic representations by enforcing invariance across augmented views. The learned representations are then transferred to downstream tasks through a separate fine-tuning stage. While effective for general-purpose representation learning, this two-stage pipeline implicitly assumes that downstream models can fully adapt to the pretrained representations, which may not hold for time series models whose inductive biases are closely tied to temporal dynamics and task-specific objectives. In contrast, AutoDA follows a one-stage, task-aware optimization scheme, where augmentation strategies are jointly optimized with the downstream task. Instead of relying on fixed or manually designed transformations, AutoDA dynamically adjusts both the selection probability and transformation intensity during training. This joint optimization produces high-quality and diverse augmented samples that are explicitly tailored to the downstream objective, directly enhancing model performance across tasks such as classification, forecasting, regression, and anomaly detection.

AutoDA-Timeseries Training Framework

MY ALT TEXT

A time series feature-aware augmented data generator is composed of multiple stacked Augmentation Layers, each of which is responsible for selecting and applying one of the available transformations. The k-th augmentation layer generates an augmentation policy consisting of (i) a series of probability indicating the likelihood of choosing transformation and (ii) a series of intensity to apply a chosen transformation. By stacking these augmentation layers, the framework can explore a variety of transformation sequences, allowing for more diverse and potentially useful augmented data. The final output augmented data is used to train a single downstream model in a single-stage, end-to-end manner, with a composite loss to update the parameters in the augmented data generator together with the downstream model.

Overall Comparison of AutoDA-Timeseries with Baselines

MY ALT TEXT

We compare AutoDA-Timeseries with three groups of baselines to ensure a comprehensive and fair evaluation. We use NoAug as the control group, which does not apply any augmentation. For representation learning, we adopt InfoTS, AutoTCL, and TS2Vec, which leverage data augmentation to construct contrastive views and learn task-agnostic representations in a two-stage manner. For automated data augmentation, we consider four state-of-the-art methods: RandAugment, UniformAugment, TrivialAugment, and A2Aug. The figure presents an overall comparison of AutoDA-Timeseries with the baselines across five time series tasks. We observe that AutoDA-Timeseries consistently achieves the best performance, covering the largest area in the radar plot.

Generalization Analysis

MY ALT TEXT

To further examine the generalizability of AutoDA-Timeseries, we conduct transfer experiments across datasets. Specifically, we train the downstream model together with augmentation policies on ETTh1 and directly evaluate the trained model on ETTh2 and ETTm2, comparing with NoAug and UniformAugment baselines. As shown in the upper block (ETTh1 $\to$ ETTh2), AutoDA-Timeseries consistently outperforms the baselines across all forecasting horizons, achieving the lowest average MSE and MAE, which demonstrates that the models trained with our framework generalize well to datasets with similar distribution. In the setting of ETTh1 $\to$ ETTm2, where the source and target distributions differ substantially, the performance gap narrows, yet AutoDA-Timeseries remains competitive and clearly superior to UniformAugment. These results highlight that AutoDA-Timeseries not only enhances performance within a single dataset but also exhibits strong potential for cross-dataset generalization, validating its robustness and applicability in real-world scenarios.

BibTeX

@inproceedings{douautoda,
        title={AutoDA-Timeseries: Automated Data Augmentation for Time Series},
        author={Dou, Zijun and Yao, Zhenhe and Xie, Zhe and Wen, Xidao and Xiao, Tong and Pei, Dan},
        booktitle={The Fourteenth International Conference on Learning Representations}
      }