Learned Data Transformation: A Data-centric Plugin for Enhancing Time Series Forecasting
Keywords: time series, data-centric, data transformation, forecasting, generalization, deep learning
Abstract: Data-centric approaches in Time Series Forecasting (TSF) often involve heuristic-based operations on data. This paper proposes to find a general end-to-end data transformation that serves as a plugin to enhance any arbitrary TSF model's performance. Our idea is to generate transformed data during an approximating process and to co-train a predictor for evaluating data with the transformation. To achieve this, we propose the Proximal Transformation Network (\model{}), which learns effective transformations while maintaining proximity to the raw data to ensure fidelity. When orthogonally integrated with popular TSF models, our method helps achieve state-of-the-art performance on seven real-world datasets. Additionally, we show that the proximal transformation process can be interpreted in terms of predictability and distribution alignment among channels, highlighting the potential of data-centric methods for future research. Our code is available at \href{https://anonymous.4open.science/r/PTN-2FC6/}{https://anonymous.4open.science/r/PTN-2FC6/}.
Primary Area: learning on time series and dynamical systems
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 10867
Loading