SDSC:A Structure-Aware Metric for Semantic Signal Representation Learning
Jeyoung Lee, Hochul Kang
TL;DR
This paper introduces SDSC, a structure-aware metric for time series representation learning that improves semantic alignment by focusing on signal structure, outperforming traditional MSE-based methods in various benchmarks.
Contribution
The paper proposes SDSC, a novel structure-aware similarity metric for signals, and demonstrates its effectiveness as a loss function in self-supervised learning tasks.
Findings
SDSC achieves comparable or better performance than MSE in forecasting and classification.
SDSC improves semantic alignment in low-resource and in-domain scenarios.
Using SDSC enhances the interpretability and structural fidelity of signal representations.
Abstract
We propose the Signal Dice Similarity Coefficient (SDSC), a structure-aware metric function for time series self-supervised representation learning. Most Self-Supervised Learning (SSL) methods for signals commonly adopt distance-based objectives such as mean squared error (MSE), which are sensitive to amplitude, invariant to waveform polarity, and unbounded in scale. These properties hinder semantic alignment and reduce interpretability. SDSC addresses this by quantifying structural agreement between temporal signals based on the intersection of signed amplitudes, derived from the Dice Similarity Coefficient (DSC).Although SDSC is defined as a structure-aware metric, it can be used as a loss by subtracting from 1 and applying a differentiable approximation of the Heaviside function for gradient-based optimization. A hybrid loss formulation is also proposed to combine SDSC with MSE,…
Peer Reviews
Decision·Submitted to ICLR 2026
1. Adapting the DSC from the segmentation domain to time-series signals is a novel perspective. Using the signed amplitude intersection as a proxy for waveform structure similarity is an interesting idea. 2. The O(T) linear complexity of SDSC is computationally efficient, which is a practical advantage. 3. The mathematical definition is intuitive, and the experimental design is well-structured.
1. A motivation for the paper is that SDSC serves as a lightweight alternative to O(T^2) metrics (e.g., SoftDTW, DILATE). However, a direct comparison against them is missing. Currently, we only know that SDSC is faster, but we do not know how much performance is lost (or gained) compared to SoftDTW. 2. The α parameter in the Sigmoid function significantly influences the gradient shape. The paper lacks a sensitivity analysis on how α affects the performance of downstream tasks. 3. The "alignment
1. The proposed method is efficient, linear time complexity, much better than other methods like DTW, which performs similar structure-awareness measurement. 2. The metric is bounded from 0 to 1, which provides better interpretability. 3. The paper is clean written, with illustrative examples to show numbers using different metric, under different structure changes.
1. No clear definition of "structure", still related to alignment or warping. 2. The backbone model is not widely tested. With more powerful models, we don't know if the advantage of SDSC still exists. 3. The imrpovement on various tasks, are very marginal. For example, in the fine-tuned classification task, the SDSC approach is not showing better results in either in-domain or out-domain experiments.
1. The paper is relatively simple and easy to understand. 2. Experiments span multiple tasks (forecasting, in-domain classification, cross-domain classification), settings (frozen encoders, fine-tuning), and datasets, demonstrating broad applicability and providing nuanced insights into when each approach works best. 3. The paper acknowledges that SDSC models achieve higher structural alignment at the cost of increased MSE, and that dataset characteristics influence which approach works better
1. The paper uses only SimMTM as the backbone "for architectural simplicity," which severely limits generalizability claims. Without validation on diverse architectures (Transformers, CNNs, RNNs, recent foundation models), it's unclear if SDSC benefits are architecture-specific or truly general. 2. The paper only compares against MSE, PCC, and SI-SNR. Recent structure-aware losses for time series (e.g., shape-based losses, spectral losses, contrastive losses) are not included, making it diffic
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsNatural Language Processing Techniques · Music and Audio Processing · Speech Recognition and Synthesis