Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReAugment: Model Zoo-Guided RL for Few-Shot Time Series Augmentation and Forecasting

Haochen Yuan, Yutong Wang, Yihong Chen, Yunbo Wang, Xiaokang Yang

TL;DR

ReAugment addresses data scarcity in time-series forecasting by using a forecasting model zoo to identify overfit-prone samples and employing a variational masked autoencoder (VMAE) as an RL actor to generate diverse, anchor-centered augmented data. The RL policy, optimized via REINFORCE, leverages a reward grounded in backtested model-zoo errors to steer augmentation toward regions where models overfit while preserving proximity to the original distribution. Key contributions include defining model-zoo variance as an overfit-prone data signal, integrating a VMAE-based augmentation backbone with RL, and demonstrating consistent improvements across five real-world datasets for both few-shot and standard forecasting tasks. The approach offers a closed-loop, task-aware augmentation paradigm with manageable computational costs, though it relies on multiple pretrained models and backtesting across the model zoo.

Abstract

Time series forecasting, particularly in few-shot learning scenarios, is challenging due to the limited availability of high-quality training data. To address this, we present a pilot study on using reinforcement learning (RL) for time series data augmentation. Our method, ReAugment, tackles three critical questions: which parts of the training set should be augmented, how the augmentation should be performed, and what advantages RL brings to the process. Specifically, our approach maintains a forecasting model zoo, and by measuring prediction diversity across the models, we identify samples with higher probabilities for overfitting and use them as the anchor points for augmentation. Leveraging RL, our method adaptively transforms the overfit-prone samples into new data that not only enhances training set diversity but also directs the augmented data to target regions where the forecasting models are prone to overfitting. We validate the effectiveness of ReAugment across a wide range of base models, showing its advantages in both standard time series forecasting and few-shot learning tasks.

ReAugment: Model Zoo-Guided RL for Few-Shot Time Series Augmentation and Forecasting

TL;DR

ReAugment addresses data scarcity in time-series forecasting by using a forecasting model zoo to identify overfit-prone samples and employing a variational masked autoencoder (VMAE) as an RL actor to generate diverse, anchor-centered augmented data. The RL policy, optimized via REINFORCE, leverages a reward grounded in backtested model-zoo errors to steer augmentation toward regions where models overfit while preserving proximity to the original distribution. Key contributions include defining model-zoo variance as an overfit-prone data signal, integrating a VMAE-based augmentation backbone with RL, and demonstrating consistent improvements across five real-world datasets for both few-shot and standard forecasting tasks. The approach offers a closed-loop, task-aware augmentation paradigm with manageable computational costs, though it relies on multiple pretrained models and backtesting across the model zoo.

Abstract

Time series forecasting, particularly in few-shot learning scenarios, is challenging due to the limited availability of high-quality training data. To address this, we present a pilot study on using reinforcement learning (RL) for time series data augmentation. Our method, ReAugment, tackles three critical questions: which parts of the training set should be augmented, how the augmentation should be performed, and what advantages RL brings to the process. Specifically, our approach maintains a forecasting model zoo, and by measuring prediction diversity across the models, we identify samples with higher probabilities for overfitting and use them as the anchor points for augmentation. Leveraging RL, our method adaptively transforms the overfit-prone samples into new data that not only enhances training set diversity but also directs the augmented data to target regions where the forecasting models are prone to overfitting. We validate the effectiveness of ReAugment across a wide range of base models, showing its advantages in both standard time series forecasting and few-shot learning tasks.
Paper Structure (33 sections, 4 equations, 3 figures, 14 tables, 1 algorithm)

This paper contains 33 sections, 4 equations, 3 figures, 14 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Transformer-based time series forecasting.
  4. Few-shot time series forecasting.
  5. Time series augmentation.
  6. Overfit-Prone Samples as Anchor Data
  7. Finding Overfit-Prone Data with a Model Zoo
  8. Preliminary Findings on Overfit-Prone Data
  9. ReAugment
  10. Overall Training Pipeline
  11. Variational Masked Autoencoder as the RL Actor
  12. Variance-Guided REINFORCE Using Model Zoo Predictions
  13. Motivation:
  14. Experiments
  15. Experimental Setups
  16. ...and 18 more sections

Figures (3)

  • Figure 1: ReAugment enables closed-loop optimization of time series augmentation and forecasting, presenting an early study on using reinforcement learning for (few-shot) time series augmentation.
  • Figure 2: Preliminary findings on overfit-prone data. We compare the performance of iTransformer trained with different splits of the original training set, which are divided based on the variance of prediction errors across the forecasting model zoo.
  • Figure 3: Architecture of ReAugment. Left: ReAugment ReAugment pretrains a VMAE as the augmentation backbone, modeling the original distribution of overfit-prone data. Right: An RL framework finetunes the VMAE prior network, using its latent space as the action space, guided by a reward function that promotes diverse sample generation around overfit-prone regions.