VIPER: Vibrant Period Representation for Robust and Efficient Time Series Forecasting

Guoqi Yu, Muxuan Li, Xiaowei Hu, Angelica I Aviles-Rivero, Shujun Wang

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: long-term forecasting, deep learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We introduce VIPER, which effectively and dynamically harnesses the inherent multi-periodic nature of time series data.

Abstract: In a data-driven world teeming with vast volumes of time series data, forecasting models play a pivotal role. The real-world time series data often exhibits intricate periodic patterns and trends, posing challenges for accurate modeling. Existing methods, reliant on fixed parameters and sampling techniques, may struggle to capture these complexities effectively. This paper designs a Vibrant Period Representation Enrichment (VIPER) framework, which effectively and dynamically harnesses the inherent multi-periodic nature of time series data. The VIPER framework adeptly separates the input sequence into trend and seasonal components. A Temporal Aggregation Block is specifically deployed for processing the seasonal component, applying innovative multi-period transformations compounded with global self-attention mechanism. This configuration enables a comprehensive capture of both short-term and long-term period information, culminating in a vibrant period representation true to the essence of the temporal dynamics. Remarkably, experimental results from eight different time series forecasting datasets substantiate the superior performance, simplicity, and computational efficiency of VIPER compared with the state-of-the-arts.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1861