Technical Summary: Signal Dice Coefficient (SDSC) for Time-Series Representation Learning

1. Problem Statement

Self-supervised learning (SSL) for time-series data has largely relied on reconstruction-based objectives using distance-based losses, most commonly Mean Squared Error (MSE). While effective at minimizing pointwise amplitude differences, the authors argue that these metrics exhibit critical limitations for semantic representation learning:

• Amplitude Sensitivity vs. Polarity Invariance: Distance-based metrics are highly sensitive to amplitude scale but largely invariant to waveform polarity. Consequently, signals that are phase-inverted, amplitude-scaled, or shifted to near-zero baselines can yield low error scores despite representing semantically distinct or reversed structures.
• Lack of Structural Fidelity: Low reconstruction error does not necessarily correlate with high semantic fidelity. For instance, a constant zero signal and a scaled waveform can produce identical MSE scores despite stark structural differences.
• Unbounded Nature: Distance-based losses are unbounded and lack normalization, complicating interpretation and model selection across different domains.

The paper posits that exclusive reliance on amplitude-centric metrics may lead to semantically incorrect reconstructions, particularly in domains like EEG or ECG where local waveform shape and polarity carry diagnostic significance.

2. Methodology

2.1. Signal Dice Similarity Coefficient (SDSC)

The authors introduce the Signal Dice Similarity Coefficient (SDSC), a structure-aware metric designed to quantify local structural agreement between temporal signals.

• Conceptual Basis: SDSC extends the Dice Similarity Coefficient (DSC), traditionally used for set overlap in semantic segmentation, to continuous, signed time-series data.
• Definition: Instead of measuring set membership, SDSC computes structural alignment based on signed amplitude intersections at each time step. It interprets the area under the curve as a proxy for waveform shape.
  • Let $E(t)$ be the ground truth and $R(t)$ be the reconstruction.
  • The intersection term $S(t)$ is defined as $E(t) \cdot R(t)$.
  • The union-like term $M(t)$ is defined as $|E(t)| + |R(t)| - \frac{||E(t)| - |R(t)||}{2}$.
  • The continuous SDSC is formulated as:
    $$SDSC(E(t), R(t)) = \frac{2 \int H(S(t)) \cdot M(t) \, dt}{\int (|E(t)| + |R(t)|) \, dt}$$
    where $H(\cdot)$ is the Heaviside step function.
• Properties: SDSC is bounded within $[0, 1]$, exhibits reduced sensitivity to amplitude variations, and explicitly reflects agreement in local waveform structure (sign and magnitude overlap). It is alignment-free and does not account for global temporal shifts or warping.

2.2. Differentiable Approximation and Training

Since the Heaviside function is non-differentiable, the authors employ a smooth, sigmoid-based approximation $\hat{H}(x) = \frac{1}{1 + e^{-\alpha x}}$ to enable gradient-based optimization. The training objective is defined as minimizing $1 - SDSC$.

2.3. Hybrid Loss Function

Recognizing that SDSC captures structure but ignores amplitude, while MSE captures amplitude but misses structure, the authors propose a Hybrid Loss:
$$L_{hybrid} = \lambda_{sdsc} \cdot L_{sdsc} + \lambda_{mse} \cdot L_{MSE}$$
The weights $\lambda_{sdsc}$ and $\lambda_{mse}$ are determined using an uncertainty-based tuning strategy (Kendall et al., 2018), where the coefficients are learned parameters representing the homoscedastic uncertainty of each loss term.

2.4. Experimental Framework

The proposed metrics are integrated into SimMTM, a self-supervised framework combining contrastive learning (InfoNCE) and reconstruction. To isolate the effect of the reconstruction objective, the contrastive component remains fixed, and only the reconstruction loss is varied (MSE, SDSC, or Hybrid).

3. Key Contributions

1. SDSC Metric: The introduction of a bounded, normalized metric that explicitly measures local structural agreement in signed time-series signals, addressing the polarity and scale insensitivity of MSE.
2. Differentiable Formulation: A practical, differentiable approximation of SDSC using a sigmoid function, allowing it to serve as a direct training objective.
3. Hybrid Objective: A novel loss function that balances structural consistency (SDSC) and amplitude fidelity (MSE) via learnable uncertainty weights.
4. Empirical Analysis: A comprehensive evaluation demonstrating that SDSC-based pretraining yields downstream performance comparable to, and in some regimes more consistent than, MSE-based objectives, particularly in in-domain and low-resource settings.

4. Experimental Results

4.1. Pre-training Analysis

• Metric Correlation: A weak negative correlation ($-0.324$) was observed between MSE and SDSC during pre-training, indicating that minimizing amplitude error does not guarantee structural alignment.
• Structural Consistency: Under fixed MSE conditions, models pre-trained with SDSC achieved significantly higher and more consistent SDSC scores compared to MSE-based models.

4.2. Downstream Forecasting

• Performance: SDSC-based models achieved forecasting performance comparable to MSE-based models on datasets like Electricity, despite exhibiting higher reconstruction error (MSE) during pre-training.
• Implication: This suggests that beyond a certain threshold, further reduction in amplitude error yields diminishing returns, and preserving local waveform structure is sufficient for effective forecasting representations.

4.3. Downstream Classification

• In-Domain: In frozen-encoder scenarios, SDSC-based pretraining consistently outperformed MSE-based pretraining in in-domain settings, highlighting its ability to preserve semantic information relevant to discrimination.
• Cross-Domain: Performance varied by dataset characteristics; amplitude-dependent features favored MSE, while structure-dependent features favored SDSC.
• Fine-Tuning: While fine-tuning narrowed performance gaps, SDSC-based models consistently demonstrated higher precision, suggesting more semantically coherent representations.

4.4. Computational Efficiency

Unlike alignment-based objectives (e.g., SoftDTW, DILATE) which suffer from quadratic computational complexity, SDSC is computationally linear and alignment-free, offering a scalable alternative.

5. Significance and Claims

The paper claims that SDSC serves as a complementary alternative to purely distance-based objectives for time-series self-supervised learning.

• Re-evaluating MSE: The results challenge the default reliance on MSE, suggesting that low reconstruction error can mask semantically inconsistent signals. SDSC exposes these limitations by explicitly penalizing structural misalignment.
• Structure-Aware Learning: The findings indicate that explicitly encouraging structural consistency in reconstruction complements distance-based objectives.
• Practicality: SDSC offers a lightweight, bounded, and interpretable metric that facilitates standardized evaluation across signal domains without the computational overhead of dynamic time warping.

The authors conclude that while SDSC does not universally outperform MSE in all downstream tasks, it provides a robust mechanism for preserving local waveform structure, which is critical for semantic representation learning, particularly in scenarios where amplitude scaling or polarity shifts are irrelevant to the task semantics.