TSGym: Automatic Model Design Framework for Deep Multivariate Time-Series Forecasting

Shuang Liang, Chaochuan Hou, Xu Yao, Shiping wang, Minqi Jiang, Songqiao Han, Hailiang Huang

Keywords: Time Series Forecasting, Large-scale Benchmark, Model Selection, Automated Machine Learning

Abstract:

Recently, deep learning has driven significant advancements in multivariate time series forecasting (MTSF) tasks.
Prevailing paradigm in MTSF research involves proposing models as pre-defined, holistic architectures. Such an approach limits adaptability across diverse data scenarios, and obscures the individual contributions of their core components.
To address this, we propose TSGym, a novel framework for automated MTSF model design. The framework begins with
decoupling existing deep MTSF methods into fine-grained components, which enables a large-scale, component-level evaluation that offers crucial insights, and creates a vast space for the automated construction of potentially superior models. Leveraging this space through strategic sampling, a core meta-learner is trained to learn the mapping between component configurations and performance across multiple traininig datasets. This enables it to perform zero-shot selection of a top-performing model for any new, unseen time series data.
Extensive experiments indicate that the model automatically constructed by our proposed TSGym significantly outperforms existing state-of-the-art MTSF methods and AutoML solutions, and exhibit high potential for
transferability across diverse datasets.

Primary Area: learning on time series and dynamical systems

Submission Number: 22920