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

• Unifying time and frequency domains is not a novel idea (e.g., [Woo et al., 2022]).

• In Introduction,

i. For decoupling the periodicity characteristics from time series data, the time-frequency transform is used for better extraction of time-frequency variation patterns in a higher dimensional feature space. ii. The mutagenicity disrupt the stability of the semantic representation space for time series. To address such issue, a group of parallel multi-scale convolution blocks is designed to deeply explore the transient patterns. iii. To capture the trend patterns, the order-preserving is added to the loss function. This learning strategy, guided by the It is not clear that the proposed methods address the corresponding problems respectively. It is not clear that the proposed methods address the corresponding problems.

• The code for reproducibility is not open.

• It is unclear if the experimental results are statistically significant, and thus they are not convincing.

• Experimental setting is not clearly stated.

• SpecAR-Net seems like a straightforward combination of STFT and multi-scale convolutions for time-series data. Limited insights/justifications are provided for the assumed modeling choices.
• The proposed approach consists of multiple complex components (short-term Fourier transform (STFT) followed by multi-scale 2D convolutions, pooling, and temporal order preserving constraint). The paper does not provide justification for model choices, e.g.,

1. STFT vs. Wavelet transforms: Unlike the STFT which uses a fixed size window, the Wavelet Transform uses a variable size window to analyze the signal, allowing for better time resolution at high frequencies and better frequency resolution at low frequencies.
2. While the paper opts for the Hamming window, alternative windows could be considered, e.g., rectangular window (no windowing), Hann window, Blackman window, Gaussian window, etc. Each of these has different properties and is suitable for different situations. It's often beneficial to experiment with different windows to see which one works best for a specific application.

• Several works jointly modeling time and frequencies of 1D time-series have been proposed, it's unclear what differentiates SpecAR-net from such approaches, e.g., TimesNet.

Minor

• The font size of the following figures should be increased for clarity: Figures 1, 3

1. It is claimed that the proposed SpecAR-Net is a plug-and-play time series representation module. Unfortunately, this claimed is fully supported. For example, the plugging of the proposed structure in other models is not studied and reported.
2. As the frequency domain based methods have been applied widely in time series, such as TimesNet and Fedformer. In fact, the proposed framework is very similar to TimesNet. It seems that the primary distinction between proposed methods and TimeNnet is the modification of convolutions into the frequency domain. Are there any other differences that contribute to the advancements in downstream tasks?
3. More analysis experiments are welcome to make the empirical studies convincing. For example, I would recommend add the experiment with Wavelet transform. What if the STFT is replaced by Wavelet Transform? It would be beneficial to add such experiments.
4. The writing can be further improved. For example, the first several sentences of Abstract are not directly related to the contribution of this paper. The resolution of Figure 1 and Figure 2 is too low to check the details.