Time2Image: A Unified Image Representation Framework for Time Series Classification
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: Time series classification; Time series image representation; Adaptive time series gaussian mapping; Vision Transformer
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Time Series Classification (TSC) is a crucial and challenging task that holds significant importance across various domains, of which one of the kernel ingredients is to construct a suitable time series representation for better feature capture. However, extracting informative and robust time series representation with good generalization potential is still a challenging problem. To address this issue, we propose Time2Image, a novel image-based representation framework for TSC. At the heart of our framework is a proposed Adaptive Time Series Gaussian Mapping (ATSGM) module for robust time series encoding in 2D image structure, based on which we employ Vision Transformer (ViT) for subsequent classification tasks considering its prominent long-dependency modeling capability. Experiments were conducted on all 158 public time series datasets from UCR/UEA covering diverse domains, among which our method achieves top 1 performance in 86 datasets compared with existing State-Of-The-Art (SOTA) methods. In addition, our framework flexibly allows handling both univariate and multivariate time series with unequal length across different domains and takes inherent advantage of generalization ability due to our proposed ATSGM representation method. The source code will be publicly available soon.
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: 9293
Loading