Signal Dice Similarity Coefficient (SDSC)

• Signal Dice Similarity Coefficient is a metric that generalizes the classical Dice coefficient to compare continuous signed signals and sparse binary support, focusing on structural overlap.
• It employs a structure-aware overlap operation—using sign consistency and minimum amplitude—to emphasize waveform shape, polarity, and temporal alignment while discounting raw amplitude errors.
• SDSC is integrated into self-supervised learning frameworks and large-scale sparse data analysis, offering improved training objectives and computational efficiency over traditional metrics.

Searching arXiv for the cited SDSC and related Dice-extension papers to ground the article in current records. Signal Dice Similarity Coefficient (SDSC) denotes a Dice-derived similarity formulation for signals rather than sets. In contemporary time-series self-supervised learning, SDSC is defined as a structure-aware metric for continuous signed temporal signals, with similarity determined by the intersection of signed amplitudes so that waveform shape, polarity, and temporal alignment are emphasized more than raw amplitude error (Lee et al., 19 Jul 2025). In a separate sparse-vector literature, the same designation is used for the classical Dice formula applied to binary signal-support sets in very high-dimensional spaces (Zadeh et al., 2012). A closely related precursor is the continuous Dice coefficient (cDC), which extends Dice from binary masks to probabilistic outputs and provides a threshold-free overlap construction that is explicitly described as a Dice-type similarity on real-valued signals (Shamir et al., 2019).

1. Terminological scope and lineage

The classical Dice Similarity Coefficient (DSC) is the common starting point for SDSC and related constructions. For sets $X$ and $Y$, it is

$DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$

In the sparse-vector setting, this same form is used on binary indicator vectors, where similarity is computed from co-occurrence counts. In the temporal-signal setting, the set intersection is replaced by overlap of signed amplitudes. In probabilistic segmentation, cDC replaces hard set cardinalities with continuous-valued overlap while preserving the Dice normalization logic (Zadeh et al., 2012, Shamir et al., 2019, Lee et al., 19 Jul 2025).

Formulation	Inputs	Core expression
DSC	Sets or binary masks	$\frac{2|X \cap Y|}{|X| + |Y|}$
cDC	Binary $A$, probabilistic $B \in [0,1]$	$\frac{2|A \cap B|}{c|A| + |B|}$
SDSC	Continuous signed signals $E(t), R(t)$	$$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$

A recurrent source of confusion is that SDSC is not a single historically fixed formula across all subfields. In one usage, it is exactly Dice on sparse binary activation patterns; in another, it is a continuous signed-signal overlap for time-series representation learning. The cDC formulation is not itself named SDSC, but it is explicitly presented as essentially a Dice-type similarity defined on real-valued signals. This suggests a broader Dice family in which overlap is generalized from sets to signal-valued objects, with different normalizations chosen for different data modalities (Shamir et al., 2019).

2. Continuous signed-signal formulation

For temporal signals, the 2025 SDSC formulation takes a ground-truth signal $E(t)$, a reconstructed or predicted signal $Y$0, and a time domain $Y$1. It introduces a sign-consistency term

$Y$2

so that $Y$3 when the two signals have the same sign and $Y$4 when they have opposite signs or one is zero. It also defines a pointwise intersection amplitude

$Y$5

which is stated to be equivalent to

$Y$6

SDSC is then defined as

$Y$7

The numerator therefore accumulates overlapping signed amplitude only where signs agree, while the denominator plays the role of the combined signal “size,” analogous to $Y$8 in DSC (Lee et al., 19 Jul 2025).

For sampled signals, the paper gives a discrete approximation: $Y$9 This form is computationally tractable for arbitrary time series and preserves the intended interpretation: similarity is the fraction of total signal mass that overlaps structurally in sign and magnitude. The paper characterizes SDSC as bounded in $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$0, with value $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$1 for perfect structural alignment and $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$2 for no overlapping signed amplitude; it is also stated to be symmetric, $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$3 (Lee et al., 19 Jul 2025).

The construction is deliberately structure-aware. A fully inverted waveform yields no contribution from regions with sign disagreement, because $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$4 when $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$5. Pure scaling affects both numerator and denominator through the minimum-overlap mechanism and the normalization term, which makes the score less sensitive to amplitude scale than MSE. Phase shifts and local misalignments are not ignored: they reduce overlap where peaks and troughs cease to coincide in sign and magnitude (Lee et al., 19 Jul 2025).

3. Relation to Dice family and continuous Dice

The relation between SDSC and earlier Dice extensions is clearest through cDC. In the segmentation setting, a binary ground-truth mask $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$6 is compared directly with a probabilistic map $DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$7 by defining

$DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$8

with

$DSC = \frac{2|X \cap Y|}{|X| + |Y|}.$9

When there is no overlap, the convention is $\frac{2|X \cap Y|}{|X| + |Y|}$0, which yields $\frac{2|X \cap Y|}{|X| + |Y|}$1. When $\frac{2|X \cap Y|}{|X| + |Y|}$2 is binary, $\frac{2|X \cap Y|}{|X| + |Y|}$3, so cDC reduces exactly to the classical Dice coefficient. The construction is proved to satisfy $\frac{2|X \cap Y|}{|X| + |Y|}$4, with $\frac{2|X \cap Y|}{|X| + |Y|}$5 if and only if the overlap is complete, and it is shown to be monotonically decreasing with the amount of overlap for equal total probability mass (Shamir et al., 2019).

cDC was introduced to address specific deficiencies of binary Dice for probabilistic outputs. The paper identifies threshold dependence, loss of probabilistic information, and the harsh treatment of partial volume effects near boundaries. cDC avoids thresholding, weights segmentation errors according to their confidence or probability, and is less biased to structure size. In simulations, thalamus produced $\frac{2|X \cap Y|}{|X| + |Y|}$6 with SD $\frac{2|X \cap Y|}{|X| + |Y|}$7 and $\frac{2|X \cap Y|}{|X| + |Y|}$8 with SD $\frac{2|X \cap Y|}{|X| + |Y|}$9, whereas the smaller subthalamic nucleus produced $A$0 with SD $A$1 and $A$2 with SD $A$3. For automatic STN segmentation, the reported scores were $A$4 and $A$5 (Shamir et al., 2019).

The conceptual link to SDSC is explicit in the cDC discussion. cDC is described as essentially a Dice-type similarity defined on real-valued signals, with overlap computed through inner-product-like terms rather than set intersection alone. The paper then presents a generic SDSC template for nonnegative signals,

$A$6

with

$A$7

which reduces exactly to cDC when $A$8 is binary and $A$9. This suggests that SDSC and cDC belong to the same design pattern: normalized continuous overlap with a normalization chosen so that perfect alignment yields similarity $B \in [0,1]$0 without hard thresholding (Shamir et al., 2019).

4. Optimization in self-supervised representation learning

Although SDSC is introduced as a similarity metric, the time-series formulation is also used as a training objective by defining

$B \in [0,1]$1

The principal technical difficulty is the Heaviside function $B \in [0,1]$2, which is non-differentiable. To support gradient-based optimization, the paper replaces it with a sigmoid-based approximation,

$B \in [0,1]$3

with derivative

$B \in [0,1]$4

Here $B \in [0,1]$5 controls sharpness: larger $B \in [0,1]$6 yields a closer approximation to the true Heaviside, but also sharper transitions and potential instability (Lee et al., 19 Jul 2025).

The same work also proposes a hybrid objective,

$B \in [0,1]$7

to balance structural fidelity and amplitude fidelity. The SDSC term is intended to encourage shape and polarity alignment, while the MSE term preserves amplitude where necessary and discourages pathological rescaling. The reported weighting strategy uses uncertainty-based weighting from Kendall et al. to adapt $B \in [0,1]$8 and $B \in [0,1]$9 according to homoscedastic uncertainty rather than relying purely on fixed manual tuning (Lee et al., 19 Jul 2025).

The experimental SSL setting is based on SimMTM, a masked time-series pre-training framework. Pre-training reconstructs masked segments of time series under three loss variants: pure MSE, pure SDSC-based loss with smooth Heaviside, and the hybrid loss. The datasets span forecasting benchmarks—ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, and Traffic—and classification benchmarks including Epilepsy and SleepEEG, with additional cross-domain target sets FD-B, Gesture, and EMG (Lee et al., 19 Jul 2025).

5. Empirical behavior and semantic interpretation

The empirical rationale for SDSC is that distance-based losses can score semantically poor reconstructions surprisingly well. The paper’s synthetic examples make this point numerically. For an inverted signal, distance-based metrics are reported as small, with $\frac{2|A \cap B|}{c|A| + |B|}$0, while SDSC gives $\frac{2|A \cap B|}{c|A| + |B|}$1, marking complete polarity reversal as structurally dissimilar. For $\frac{2|A \cap B|}{c|A| + |B|}$2 and $\frac{2|A \cap B|}{c|A| + |B|}$3 scaled signals, SDSC gives the same value, $\frac{2|A \cap B|}{c|A| + |B|}$4, despite strongly differing MSE and MAE. A zero signal and a $\frac{2|A \cap B|}{c|A| + |B|}$5 scaled waveform both yield $\frac{2|A \cap B|}{c|A| + |B|}$6, yet SDSC separates them as $\frac{2|A \cap B|}{c|A| + |B|}$7 versus $\frac{2|A \cap B|}{c|A| + |B|}$8. For noise-like output versus a shifted waveform, the reported SDSC values are $\frac{2|A \cap B|}{c|A| + |B|}$9 and $E(t), R(t)$0, respectively (Lee et al., 19 Jul 2025).

On pre-training diagnostics, MSE and SDSC are only weakly coupled. On ETTh1, the reported Pearson correlation between MSE and SDSC is about $E(t), R(t)$1, which the authors use to argue that decreasing MSE does not reliably increase structural similarity. Conditioning on a fixed MSE band such as $E(t), R(t)$2, SDSC-based pre-training yields higher SDSC values and lower variance with a tighter interquartile range, indicating more consistently structure-aligned reconstructions even when numerical reconstruction error is matched (Lee et al., 19 Jul 2025).

Downstream forecasting results show little separation among pre-training losses in final MSE or MAE. Across ETTh1/2, ETTm1/2, Weather, Electricity, and Traffic, pre-training with MSE, SDSC, or Hybrid leads to very similar forecasting performance, with average MSE/MAE described as essentially the same, approximately $E(t), R(t)$3-$E(t), R(t)$4 MSE and $E(t), R(t)$5 MAE. The interpretation given is that once structural aspects are adequately captured, additional pre-training MSE reduction may not translate into better forecasting (Lee et al., 19 Jul 2025).

Classification results are more discriminative. In frozen-encoder evaluation, Epilepsy$E(t), R(t)$6Epilepsy gives an average metric of $E(t), R(t)$7 for SDSC versus $E(t), R(t)$8 for MSE, and SleepEEG$E(t), R(t)$9SleepEEG gives $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$0 versus $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$1. The paper reports that SDSC and Hybrid pre-training outperform MSE in accuracy, precision, recall, F1, and their average for the in-domain Epilepsy case, while cross-domain results are mixed. Under end-to-end fine-tuning, the differences become very close, though SDSC occasionally gives slightly higher average scores on some cross-domain tasks such as SleepEEG$$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$2FD-B and SleepEEG$$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$3Gesture (Lee et al., 19 Jul 2025).

These results ground the paper’s semantic interpretation of SDSC. A high SDSC is taken to mean that peaks and troughs coincide in both sign and position, preserving characteristic waveform structure. The bounded $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$4 scale is also presented as more interpretable than unbounded MSE for cross-dataset comparison (Lee et al., 19 Jul 2025).

6. Sparse-signal SDSC and dimension-independent computation

In the DISCO literature, SDSC is used in a more classical combinatorial sense. For high-dimensional signal data represented as sparse binary activation patterns across windows, the Signal Dice Similarity Coefficient is exactly

$$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$5

where $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$6 and $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$7 are the numbers of windows in which the signals are active, and $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$8 is the number of windows in which both are active. This is the standard Dice coefficient on support sets, transferred to sparse signal vectors (Zadeh et al., 2012).

The distinctive contribution of the DISCO work is algorithmic rather than definitional. To compute all pairwise similarities between very high-dimensional sparse vectors without dependence on the ambient dimension $$\frac{ 2 \int H(S(t)) \cdot M(t) \, dt }{ \int (E(t) + R(t)) \, dt }$$9 after reading the data, the paper samples pair co-occurrences in MapReduce. For Dice similarity, the mapper emits a co-occurring pair $E(t)$0 with probability

$E(t)$1

and the reducer outputs

$E(t)$2

where $E(t)$3 is the number of sampled co-occurrences. The resulting estimator is unbiased: $E(t)$4 For Dice, the reported expected shuffle size is

$E(t)$5

and the reduce-key size is

$E(t)$6

both independent of the ambient dimension $E(t)$7 (Zadeh et al., 2012).

The paper’s experiments use Twitter-scale data and report that shuffle reduction versus the naive method is $E(t)$8 for Dice. Although the largest production-scale experiment in the paper is reported for cosine, the same complexity bounds are stated to hold for Dice, making sparse-support SDSC feasible for very large collections of high-dimensional signals so long as the per-window sparsity assumption is satisfied (Zadeh et al., 2012).

7. Limitations, misconceptions, and open directions

Several limitations follow directly from the current SDSC formulations. In the temporal-signal version, amplitude-insensitivity is deliberate, but this also means that pure SDSC may miss signal properties when exact amplitude is critical. The paper therefore treats hybrid SDSC+MSE training as preferable when amplitude still carries important information. The Heaviside-based structure term also introduces differentiability and stability issues, controlled only approximately by the sigmoid sharpness parameter $E(t)$9. Heavy noise can create frequent sign fluctuations, and the additional non-linear operations impose modest computational overhead relative to plain MSE (Lee et al., 19 Jul 2025).

A second misconception is that SDSC simply replaces MSE with a normalized score. The reported examples show a more specific behavior: polarity reversal can collapse similarity to $Y$00, equal structural distortions under different scalings can receive the same SDSC, and phase shifts retain only partial credit. SDSC is therefore not amplitude-blind in a trivial sense; rather, it is designed to weight structural agreement over pointwise magnitude error (Lee et al., 19 Jul 2025).

A third issue is terminological. The sparse-support SDSC of the DISCO framework and the continuous signed-signal SDSC of time-series SSL are not identical objects, even though both inherit the Dice normalization. cDC adds yet another closely related construction by extending Dice to a binary-versus-probabilistic setting and showing how threshold-free overlap measures can remain bounded, monotone with overlap, and less biased to structure size. This suggests that SDSC is best understood not as a single canonical formula, but as a family of Dice-inspired signal overlap measures whose precise definition depends on whether the signal is binary, probabilistic, or continuous and signed (Zadeh et al., 2012, Shamir et al., 2019).

The future directions stated for the temporal-signal SDSC include integration into contrastive, diffusion-based, and large-scale TS-PTM frameworks; investigation in domain adaptation and cross-modal settings such as audio, biomedical, and industrial sensor data; and variants including multi-scale SDSC, phase-aware versions, and alternative smooth approximations or normalization schemes. Across these extensions, the central design principle remains constant: overlap should be defined on the signal itself, without hard thresholding, in a way that preserves the Dice intuition of normalized shared structure (Lee et al., 19 Jul 2025).