Rethinking Time Series Domain Generalization via Structure-Stratified Calibration

This paper proposes a structurally stratified calibration framework that improves time series domain generalization by explicitly distinguishing structurally compatible samples and performing amplitude calibration within these clusters to avoid the negative transfer caused by global alignment across heterogeneous dynamical systems.

The Gist

Imagine you are a chef trying to teach a robot to cook the perfect soup.

The Problem: The "One-Size-Fits-All" Mistake
Usually, when we train AI to recognize patterns in time-series data (like heartbeats, sleep cycles, or daily movements), we assume that a "heartbeat" from a hospital in New York is basically the same as a "heartbeat" from a clinic in Tokyo. We try to force the AI to treat all these different data sources as if they belong to the same single "flavor profile."

The paper argues that this is a mistake. It's like trying to teach the robot to make soup by mixing ingredients from a spicy Thai curry, a creamy French bisque, and a clear Japanese broth into one giant pot. Even if they are all "soups," their fundamental structures are different. If you force them to align perfectly, you end up with a muddy, unrecognizable mess. The AI gets confused, learns the wrong rules, and fails when it tries to cook for a new customer (a new dataset) it hasn't seen before.

The Insight: "Apples to Apples, Oranges to Oranges"
The authors realized that time-series data often comes from different "families" of underlying systems.

• Analogy: Think of time-series data as music.
  • Dataset A might be a Jazz band (complex, improvisational rhythms).
  • Dataset B might be a Marching Band (strict, steady beats).
  • Dataset C might be a Heavy Metal band (distorted, high energy).

If you try to tell the AI that a Jazz drum solo and a Marching Band snare roll are "the same thing" just because they are both drums, the AI will get confused. The structure of the sound is fundamentally different.

The Solution: The "Stratified Calibration" Framework (SSCF)
The paper proposes a new method called Structure-Stratified Calibration. Here is how it works, step-by-step:

1. The Sort (Stratification):
   Before the AI tries to learn, it first looks at the "shape" of the data. It doesn't look at the labels (like "sleep" or "awake") yet; it looks at the spectrum (the frequency and energy patterns).

   • Metaphor: Imagine you have a giant pile of musical instruments. Instead of throwing them all into one room, you first sort them into three separate rooms: Room A (Jazz), Room B (Marching), and Room C (Metal). You only put instruments that sound structurally similar in the same room.

2. The Reference (Anchors):
   Inside each room, the AI creates a "perfect average" example.

   • Metaphor: In the Jazz room, it builds a "Perfect Jazz Reference." In the Marching room, it builds a "Perfect Marching Reference."

3. The Calibration (The Fix):
   Now, when a new piece of data comes in (a new song), the AI asks: "Which room does this belong to?"

   • If it's a Jazz song, the AI compares it only to the Perfect Jazz Reference. It adjusts the volume and tone to match the Jazz style.
   • It never tries to force the Jazz song to look like the Marching song.
   • Why this helps: It prevents the AI from making "spurious correspondences" (fake connections). It stops the AI from trying to turn a drum solo into a snare roll.

The Result
By sorting the data first and only comparing "apples to apples," the AI learns much faster and more accurately. When it encounters a brand new dataset (a new city, a new sensor), it can quickly figure out which "room" that data belongs to and apply the right rules.

In Summary

• Old Way: "Everything is the same. Force them all to look alike." -> Result: Confusion and failure.
• New Way (SSCF): "First, sort them by their fundamental structure. Then, fix the small differences within each group." -> Result: A smarter, more reliable AI that works better in the real world.

The paper tested this on 19 different datasets (including sleep tracking and heart monitoring) and found that this "sort-first, fix-later" approach significantly outperformed all previous methods, especially when the AI had to work on data it had never seen before.

Technical Summary

1. Problem Statement

The paper addresses the critical bottleneck of cross-dataset generalization in time-series modeling. While models often perform well on source datasets, their performance degrades significantly when transferred to unseen target domains (zero-shot setting).

• Root Cause: Existing Domain Generalization (DG) methods (e.g., CORAL, MMD, DANN) typically assume that samples from different domains are comparably meaningful within a single shared representation space. They enforce global alignment to minimize distribution shifts.
• The Flaw: In real-world scenarios, time series often originate from structurally heterogeneous latent dynamical systems. Different datasets may exhibit fundamentally distinct spectral shapes or energy distributions due to underlying generative processes, not just acquisition noise.
• Consequence: Enforcing global alignment on structurally incompatible samples leads to spurious correspondences (matching a "triangle" to a "circle") and induces negative transfer, where the model learns incorrect relationships, ultimately harming generalization.

2. Methodology: Structure-Stratified Calibration Framework (SSCF)

The authors propose SSCF, a framework that shifts the paradigm from "global alignment" to "comparability-first, calibration-second." The core idea is to explicitly identify structurally consistent subsets before performing any calibration.

Key Components:

1. Spectral Modeling Assumption:

   • Time series are modeled as $H^{(d)}(f) \approx g_d(f) \odot Z(f)$, where $Z(f)$ is a stable, structure-related latent spectral pattern, and $g_d(f)$ represents domain-specific amplitude/scale effects.
   • The goal is to align $g_d(f)$ (amplitude) while preserving $Z(f)$ (structure).

2. Structure Stratification (Offline Phase):

   • Feature Extraction: Source domain samples are encoded into shallow feature maps.
   • Spectral Representation: Power Spectral Density (PSD) is estimated for each sample using the Welch method.
   • Clustering: Samples are clustered using K-Means in the power-spectral space to form $K$ structurally compatible strata (clusters). This partitions the data based on spectral morphology (e.g., peak locations, dominant bands) rather than semantic labels alone.

3. Reference Anchor Construction:

   • For each structural stratum $k$, a Mean-Amplitude-Squared (MAS) template is constructed.
   • This involves taking the square root of the power spectra (to get amplitude), averaging them, and squaring the result. This creates a robust reference anchor $\bar{P}_k$ that is less sensitive to outliers than simple power averaging.

4. Structural Matching & Intra-Stratum Calibration (Online Phase):

   • Matching: For any input sample (source or target), its spectral representation is computed and matched to the nearest reference anchor $\bar{P}_{k^*}$ based on Euclidean distance in the joint channel-frequency space.
   • Calibration: Amplitude calibration is performed exclusively within the matched stratum.
     • Phase Preservation: The phase information of the original signal is strictly preserved.
     • Amplitude Scaling: The amplitude is scaled to match the reference anchor $\bar{P}_{k^*}$.
     • Formula: The scaling factor $M$ is derived in closed form: $M(c, f) = \sqrt{\frac{\bar{P}_{k^*}(c, f)}{P(c, f) + \epsilon}}$.
   • This ensures that samples are only aligned with others that share similar underlying dynamical structures.

5. Training Protocol:

   • Stage 1: Train an auxiliary encoder to extract features and construct fixed structural anchors (stratification and anchors are frozen).
   • Stage 2: Train the final encoder-classifier end-to-end. The calibration operation is differentiable with respect to the encoder outputs but the anchors remain fixed constants.

3. Key Contributions

1. Theoretical Insight: The paper identifies a fundamental applicability boundary for alignment methods: global alignment is ill-posed under structural heterogeneity without assessing comparability. It argues that structural consistency must be established prior to alignment.
2. Novel Framework (SSCF): Introduces a "stratify-first, calibrate-within-stratum" approach that explicitly handles structural heterogeneity by restricting calibration to compatible subsets, thereby eliminating spurious cross-domain correspondences.
3. Empirical Validation: Demonstrates significant performance improvements across 19 public datasets (100.3k samples) covering three distinct tasks: Sleep Staging, Arrhythmia Detection, and Human Activity Recognition (HAR).

4. Experimental Results

The framework was evaluated under a Leave-One-Domain-Out (LODO) protocol and on external target domains in a zero-shot setting.

• Sleep Staging: SSCF achieved a Macro-F1 score of 69.37% on LODO and 75.12% on external targets, significantly outperforming strong baselines like CORAL (71.33% external) and IRM (66.65% external).
• Arrhythmia Detection: SSCF improved the average Macro-F1 to 81.26%, surpassing baselines which struggled with specific target domains (e.g., MMD dropped to 65.44%).
• Human Activity Recognition (HAR): SSCF achieved 94.43% Macro-F1, outperforming the next best method (IRM at 91.83%).
• Ablation Studies:
  • Global Alignment: Applying a single global anchor degraded performance, confirming that ignoring structural differences causes negative transfer.
  • Granularity ($K$): Performance stabilized when $K \ge 3$, indicating that coarse-grained stratification is sufficient; fine-grained partitioning is not necessary.
  • Matching Quality: Higher matching quality (closer geometric distance to the anchor) correlated strongly with better performance, validating the structural matching mechanism.

5. Significance

• Paradigm Shift: The paper challenges the prevailing assumption that "more alignment is better." It posits that structural compatibility is a prerequisite for meaningful alignment.
• Robustness: By preventing the mixing of incompatible spectral structures, SSCF provides a more reliable pathway for deploying time-series models in open, real-world environments where acquisition conditions and underlying dynamics vary.
• Efficiency: The method is computationally efficient, relying on simple spectral clustering and closed-form amplitude scaling, making it scalable to large datasets without complex generative modeling.
• Generalizability: The approach is domain-agnostic regarding the specific time-series task, as demonstrated by its success across medical (EEG/ECG) and physical (HAR) domains.

In conclusion, SSCF offers a principled solution to the negative transfer problem in time-series domain generalization by ensuring that calibration occurs only between samples that are structurally comparable, thereby establishing a more robust foundation for cross-domain learning.