Meta-Learning for Time Series Analysis and/or Forecasting: Concept Review and Comprehensive Critical Comparative Survey

Meta-Learning has emerged as a solution to address the limitations of data unavailability and the lack of extensive computing resources. The aim of this work is to consolidate a discussion on the application of Meta-Learning in Time Series forecasting, specifically by comprehensively contrasting Zero-Shot Learning (ZSL), One-Shot Learning (OSL), and Few-Shot Learning (FSL). In our implementation setup, the comparative analysis of results identifies Few-Shot Learning as the best performer and One-shot learning as the worst. The performance improvement registered for FSL is credited to the additional meta-knowledge learned through the MAML-based method. Implemented as a Siamese network, OSL had to learn the strong periodic components within the time series. However, the selected datasets did not display such strong periodicity; instead, they were dominated by trend and noise components. Future work intends to analyze the asymptotic performance of increasingly complex predictors in tandem with increasingly longer training regimes.