Towards plausibility in time series counterfactual explanations

This paper introduces a gradient-based optimization method for generating time series counterfactual explanations that ensures plausibility by integrating soft-DTW alignment with k-nearest neighbors, resulting in valid and temporally realistic outputs that outperform existing approaches in distributional alignment.

The Gist

Imagine you have a very smart, but somewhat mysterious, robot doctor. This robot looks at your heart rate monitor (a time series) and says, "You have a heart condition."

You ask, "Why?" The robot says, "Because of the pattern." But that's not helpful. You want to know: "What is the minimum change I need to make to my heart rhythm so that the robot would say, 'You are healthy' instead?"

This is what the paper calls a Counterfactual Explanation. It's like asking, "What if I had done X instead of Y?"

The Problem: The "Fake" Heartbeat

The authors found that existing methods for answering this question often create "fake" heartbeats.

Imagine you are trying to explain to a friend how to fix a broken clock.

• Old methods might say: "Just take the gears out and glue them back on in a random order that looks like a clock." Technically, it might tick, but it looks weird and unnatural. If you tried to build a real clock that way, it would be a disaster.
• In the world of data, these methods create "adversarial" patterns—tiny, weird glitches that trick the robot into changing its mind, but they don't look like anything a real human heart (or stock market, or factory machine) would ever actually do. They lack plausibility.

The Solution: The "Realistic" Makeover

The authors, Marcin, Krzysztof, and Maciej, built a new tool that acts like a talented stylist rather than a random glitch generator.

Here is how their method works, using a simple analogy:

1. The Goal: A Valid Change

First, the tool must change the input (your heart rate) just enough so the robot changes its prediction from "Sick" to "Healthy." This is the Validity check.

2. The Secret Sauce: The "Target Class" Dance Floor

This is the paper's big innovation. To make sure the new heart rate looks real, the tool doesn't just guess. It goes to a "dance floor" filled with thousands of examples of healthy heartbeats (the target class).

It picks the 10 closest healthy heartbeats to your current one. Think of these as your "role models."

3. The Soft-DTW: The Flexible Ruler

Usually, comparing two heartbeats is like comparing two lines of text word-for-word. If one heartbeat is slightly faster or slower than the other, a standard ruler says they are totally different.

The authors use a special tool called Soft-DTW (Dynamic Time Warping).

• Analogy: Imagine two people dancing to the same song, but one is slightly ahead of the beat. A standard ruler would say, "You are out of sync!"
• Soft-DTW is like a flexible, stretchy ruler. It says, "You are dancing the same moves, just at a slightly different speed. You are still in sync."
• The tool uses this flexible ruler to gently nudge your "sick" heartbeat until it looks like it belongs in the same dance circle as the "healthy" role models.

The Trade-off: Comfort vs. Style

The paper admits there is a catch.

• Old methods try to change as little as possible (minimal effort), even if the result looks weird.
• The new method is willing to make a bigger change to the data if it means the result looks realistic.

Analogy:
Imagine you want to turn a square peg into a round one.

• Old method: You shave off a tiny corner. It's still a square peg, but the robot thinks it's round because of a trick. It's a "cheap" fix.
• New method: You shave off more wood to make it a perfect circle. It takes more work (more "distance" from the original), but it's a real round peg that fits perfectly in the hole.

Why Does This Matter?

If you are a doctor, a financial analyst, or a factory manager, you don't want to be told, "If you change your data by this weird, impossible amount, you'll be safe." You want to know, "If I adjust my process to look more like a healthy example, will I be safe?"

This new method ensures that the "what-if" scenarios it generates are plausible. They aren't just mathematically correct; they are realistic and could actually happen in the real world.

Summary

• The Problem: Current AI explanations for time-series data often create fake, unrealistic scenarios.
• The Fix: A new method that gently reshapes the data to look like real examples from the "good" category, using a flexible comparison tool (Soft-DTW).
• The Result: Explanations that are trustworthy because they respect the natural rhythm and structure of the data, even if they require a slightly bigger change to get there.

Technical Summary

Here is a detailed technical summary of the paper "Towards plausibility in time series counterfactual explanations":

1. Problem Statement

Explainable AI (XAI) is critical for high-stakes domains like healthcare and finance where time series classifiers (e.g., for ECG analysis or fraud detection) are deployed. While Counterfactual Explanations (CFEs) offer actionable insights by showing minimal input changes required to alter a model's prediction, existing methods for time series data suffer from a lack of plausibility.

• The Gap: Current methods often generate CFEs that are mathematically valid (changing the prediction) but temporally unrealistic. They may introduce adversarial patterns, abrupt discontinuities, or structures that do not align with the natural dynamics of the target class.
• The Goal: To generate CFEs that are not only valid but also plausible, meaning they adhere to the realistic temporal structures and distributions of the target class data.

2. Methodology

The authors propose a novel gradient-based optimization method that operates directly in the input space. The core innovation is the integration of soft-DTW (Dynamic Time Warping) into the loss function to enforce plausibility.

Key Components:

• Optimization Objective: The method minimizes a multi-faceted loss function ($L_{CF}$) composed of four terms:

  1. Proximity ($L_{prox}$): Squared Euclidean distance ($L_2$) to ensure the CFE remains close to the original input.
  2. Sparsity ($L_{sparse}$): $L_1$ norm of the perturbation to encourage localized changes rather than global noise.
  3. Validity ($L_{valid}$): Hinge loss ensuring the classifier predicts the target class with high confidence.
  4. Plausibility ($L_{DTW}$): The novel component. It minimizes the soft-DTW distance between the generated CFE and the $k$-nearest neighbors (k-NN) of the target class from the training set.

• Soft-DTW Integration:

  • Standard DTW is non-differentiable, making it unsuitable for gradient descent.
  • The authors use soft-DTW, which replaces the hard minimum in DTW with a differentiable soft-minimum operator (using a smoothing parameter $\gamma$).
  • This allows the optimizer to align the generated CFE with the temporal dynamics of real target-class samples, ensuring the resulting sequence looks like a natural instance of that class.

• Process:

  1. Start with an input $X$ classified as $\hat{y}$.
  2. Initialize a candidate counterfactual $X'$.
  3. Iteratively update $X'$ via gradient descent to minimize $L_{CF}$ while keeping classifier weights fixed.
  4. The $L_{DTW}$ term pulls $X'$ toward the manifold of the target class, preventing adversarial artifacts.

3. Key Contributions

1. Novel Plausibility Mechanism: Introduction of a differentiable soft-DTW loss aligned with target-class k-NN to explicitly enforce realistic temporal structures in CFEs.
2. Comprehensive Evaluation: A rigorous benchmark against strong reference methods (Glacier and M-CELS) across eight diverse datasets (both univariate and multivariate).
3. Qualitative & Quantitative Analysis: Demonstration that existing methods often fail to preserve temporal coherence, whereas the proposed method generates CFEs that are structurally consistent with the target class, even if they require larger perturbations.

4. Experimental Results

The method was evaluated on datasets from UCI/UEA repositories (e.g., CBF, TwoLeadECG, GunPoint, Cricket) using a 1D CNN classifier.

• Validity: The proposed method achieved near-perfect validity (1.000) across almost all datasets, significantly outperforming reference methods (e.g., Glacier achieved 0.360 on CBF; M-CELS achieved 0.226).
• Plausibility (DTW Distance): The method achieved significantly lower DTW distances to target-class neighbors compared to baselines.
  • Example: On the TwoLeadECG dataset, the proposed method achieved a DTW of 0.016, compared to 0.064 (Glacier) and 0.302 (M-CELS).
  • Example: On the Cricket dataset, the proposed method scored 0.810, while M-CELS scored 65.924.
• Plausibility (Isolation Forest): The generated CFEs were frequently classified as "nominal" (non-outliers) by Isolation Forests, with perfect scores (1.000) on six datasets.
• Trade-off (Proximity vs. Plausibility): The method exhibited higher $L_1$ and $L_2$ distances (larger perturbations) compared to baselines. The authors argue this is a necessary trade-off: achieving temporal realism requires more substantial, meaningful structural changes than simply minimizing pixel-wise distance.
• Qualitative Findings: Visual analysis of ECG and CBF data showed that baselines often produced "adversarial-looking" noise or failed to capture key temporal features (e.g., specific peaks), while the proposed method successfully reconstructed the target class's geometric shape.

5. Significance and Limitations

Significance:

• The paper establishes that plausibility is a distinct and critical property for time series CFEs that cannot be achieved by simply minimizing distance metrics.
• It demonstrates that soft-DTW alignment is an effective, differentiable mechanism to guide counterfactual generation toward realistic data manifolds.
• It provides a new standard for evaluating time series CFEs, emphasizing distributional alignment over mere proximity.

Limitations:

• Computational Cost: Soft-DTW has quadratic complexity with respect to sequence length. Combined with k-NN calculations at every optimization step, the method is computationally expensive for very long time series.
• Multi-modal Distributions: The reliance on k-NN assumes the target class has consistent temporal patterns. If a class has high variability or multi-modal distributions, the k-NN alignment might force the CFE into a specific sub-pattern that does not represent the full diversity of the class.

Future Work:
The authors suggest exploring probabilistic generative models to model time series density more effectively, which could better capture complex, multi-modal temporal patterns within target classes.