SpecAR-Net: Spectrogram Analysis and Representation Network for Time Series

Y Dong, Liwen Zhang, Youcheng Zhang, Shi Peng, Wen Chen, Zhe Ma, Xuhui Huang

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 analysis, time series representation, time-frequency transformation, complex convolution

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

Abstract: Time series analysis involves modeling time series to extract valuable information, which finds broad applications in domains such as device malfunction diagnosis, human activity recognition, and medical-assisted diagnosis. Representing temporal-structured samples is crucial for time series analysis tasks. Recently, several advanced deep learning models, i.e., recurrent neural networks, convolutional neural networks, and transformer-style models, have been successively applied in the field of temporal data representation, yielding notable results. Those existing methods primarily model and represent the variation patterns within time series solely in time domain. However, as a highly abstracted information entity, time series data is formed by the coupling of various patterns such as trends, seasonality, and dramatic changes (instantaneous high dynamic), it is difficult to exploit these highly coupled properties only by means of analysis in the time domain. Consequently, it would be insufficient for time-domain dependent only methods to overcome the semantic representation bottleneck or construct comprehensive feature representations of 1D time series. To this end, we present Spectrum Analysis and Representation Network (SpecAR-Net). SpecAR-Net aims at learning more comprehensive representations by modeling raw time series in time-frequency domain, where an efficient joint extraction of time-frequency features is achieved through a group of learnable 2D multi-scale parallel complex convolution blocks. Experimental results show that the SpecAR-Net achieves excellent performance in five major downstream tasks of time series analysis i.e., classification, anomaly detection, imputation, long- and short-term series forecasting.

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: 7634