📊 statistics

Multi-SpaCE: Multi-Objective Subsequence-based Sparse Counterfactual Explanations for Multivariate Time Series Classification

This paper introduces Multi-SpaCE, a novel multi-objective counterfactual explanation method for multivariate time series classification that leverages NSGA-II to generate valid, sparse, and plausible solutions while balancing proximity and contiguity to overcome the limitations of existing univariate approaches.

Original authors: Mario Refoyo, David Luengo

Published 2026-07-20

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Mario Refoyo, David Luengo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are talking to a super-smart robot that can predict the future, like diagnosing a disease from a heartbeat or spotting a stock market crash before it happens. These robots are built using "Deep Learning," a type of artificial intelligence that is incredibly good at solving hard problems but is also a total "black box." You feed it data, and it gives you an answer, but it can't tell you why it made that choice. It's like a friend who says, "I know you'll fail this test," but refuses to explain which questions you got wrong. This is a big problem because in real life—like in hospitals or banks—we need to trust the robot's logic before we act on it.

To fix this, scientists use something called "Counterfactual Explanations." Think of this as a "What if?" game. Instead of just saying "You failed," the robot says, "You would have passed if you had studied just one more chapter." It finds the smallest, most logical change needed to flip a result from a failure to a success. But here's the tricky part: when the data isn't just a list of numbers (like a spreadsheet) but a flowing stream of information over time—like a heartbeat monitor or a weather sensor—making these "What if?" suggestions gets messy. The robot might suggest changes that are impossible in the real world, or it might change too many things at once, making the explanation confusing.

This is where a new method called Multi-SpaCE comes in. The researchers behind it, Mario Refoyo and David Luengo, realized that existing tools for time-series data were too rigid. They were like a tailor trying to fit a suit by only cutting one straight line, ignoring that the fabric has different patterns and flows. Multi-SpaCE is a smarter, more flexible tailor. It uses a clever search strategy (a genetic algorithm, which mimics how nature evolves) to find the perfect "What if?" scenarios for complex, multi-channel data.

The paper finds that Multi-SpaCE is a game-changer because it guarantees that the "What if?" scenarios it creates actually work. While other methods often suggest changes that sound good but fail to actually change the robot's mind (like suggesting you study a chapter that doesn't exist), Multi-SpaCE ensures 100% validity. It also handles data with many moving parts at once (multivariate), like a heart monitor tracking heart rate, blood pressure, and oxygen levels simultaneously, rather than just one signal.

The researchers tested their method on dozens of real-world datasets, from heartbeats to hand movements, and found that Multi-SpaCE consistently produced valid explanations that were also sparse (changing as little as possible) and plausible (looking like real data). Unlike other methods that force you to pick a single "best" answer by guessing how much you care about different factors, Multi-SpaCE gives you a whole menu of options. It presents a "Pareto front," which is like a menu where you can see the trade-offs: "Here is an explanation that changes very little but looks a bit weird," or "Here is one that looks very natural but changes a bit more." This lets the user pick the explanation that fits their specific needs, ensuring that the robot's reasoning is not only correct but also understandable and trustworthy.

Technical Summary: Multi-SpaCE

Problem Statement
Deep learning systems, while effective in complex tasks like multivariate time series classification, often lack transparency, hindering their deployment in high-stakes domains. Counterfactual explanations (CFEs) offer a solution by identifying minimal input changes required to alter a model's prediction. However, existing CFE methods for time series face three critical limitations:

Univariate Bias: Many methods are restricted to univariate data, failing to address the complexities of multivariate time series common in real-world applications (e.g., healthcare, finance).
Rigid Constraints: Approaches often enforce fixed-length subsequences or limit modifications to a single subsequence, reducing flexibility and potentially introducing unnecessary changes.
Validity Issues: Validity (ensuring the counterfactual actually changes the predicted class) is frequently treated as an optimization objective rather than a strict constraint. Consequently, many heuristic methods fail to guarantee valid explanations, rendering their other performance metrics (sparsity, plausibility) meaningless.

Methodology: Multi-SpaCE
The paper introduces Multi-SpaCE, a multi-objective counterfactual explanation method designed specifically for multivariate time series. It extends the authors' previous work, Sub-SpaCE, by integrating a multi-objective optimization framework based on the Non-Dominated Sorting Genetic Algorithm II (NSGA-II).

Core Mechanism: The method generates counterfactuals ( $x'$ ) by identifying a "Nearest Unlike Neighbor" (NUN)—the closest instance from a different class—and substituting specific subsequences of the original input ( $x$ ) with values from the NUN.
Multivariate Adaptation: To handle multivariate data, Multi-SpaCE employs a dual-mask strategy:
- Common Mask: A single binary mask shared across all channels, used initially to optimize global changes.
- Independent Mask: A separate binary mask for each channel, used in later stages to refine solutions and allow channel-specific modifications.
Mutation Operators: The genetic algorithm utilizes specialized mutation operators adapted for time series:
- Subsequence Extension/Compression: Modifies the length of existing contiguous change segments.
- Subsequence Removal: Eliminates entire subsequences to accelerate convergence in complex multivariate settings.
Objective Functions: The optimization balances four conflicting objectives without manual hyperparameter tuning:
1. Validity: Enforced strictly via a large penalization term ( $\nu$ ) for any solution that does not change the predicted class.
2. Proximity: Minimizes the distance between the original and counterfactual instances.
3. Sparsity: Minimizes the total number of changed time steps.
4. Contiguity: Minimizes the number of distinct subsequences of changes to enhance interpretability.
5. Plausibility: Penalizes deviations from the data manifold using an autoencoder-based outlier score.
Output: Instead of a single solution, Multi-SpaCE returns a Pareto front of non-dominated solutions, allowing end-users to select explanations that best fit their specific trade-off preferences between sparsity, plausibility, and contiguity.

Key Contributions

First Multi-Objective Multivariate Method: Multi-SpaCE is presented as the first method to provide a multi-objective optimization framework specifically for generating valid counterfactuals in multivariate time series classification.
Strict Validity Guarantee: Unlike many existing approaches that treat validity as a soft constraint, Multi-SpaCE enforces it as a hard requirement, ensuring 100% validity in experimental results.
Flexible Subsequence Handling: The method avoids rigid fixed-length or single-subsequence constraints, utilizing a genetic algorithm to discover optimal, variable-length subsequences across multiple channels.
Pareto-Front Generation: By leveraging NSGA-II, the method provides a diverse set of optimal solutions, eliminating the need for users to manually tune utility function weights to balance conflicting objectives.

Experimental Results
The authors evaluated Multi-SpaCE on 15 univariate and 10 multivariate datasets from the UCR and UEA archives, comparing it against state-of-the-art baselines (e.g., Native Guide, Glacier, COMTE, AB-CF, DiscoX).

Validity: Multi-SpaCE achieved perfect validity (1.0) across all datasets. In contrast, many baseline methods (particularly in multivariate settings) showed significantly lower validity rates, with some failing to produce valid counterfactuals for the majority of instances.
Proximity: While not explicitly optimized as a primary goal, Multi-SpaCE achieved competitive proximity scores, ranking second on average for univariate datasets and first for multivariate datasets among valid solutions.
Plausibility: Multi-SpaCE demonstrated robust plausibility, ranking highly across different outlier detection models (Autoencoder, Isolation Forest, Local Outlier Factor). Notably, it maintained high plausibility while other methods with better scores often suffered from low validity.
Sparsity and Contiguity: The method effectively balanced sparsity and the number of subsequences, achieving the lowest median values in combined metrics for both univariate and multivariate datasets.

Significance and Claims
The paper claims that Multi-SpaCE addresses a fundamental gap in time series explainability by ensuring that counterfactual explanations are not only interpretable and sparse but, crucially, valid. The authors argue that without guaranteed validity, other performance metrics are irrelevant. By providing a Pareto front of solutions, Multi-SpaCE offers flexibility for diverse user needs without requiring extensive experimentation to tune hyperparameters. The method is presented as a model-agnostic, computationally efficient tool that outperforms existing approaches in reliability and adaptability for real-world multivariate time series applications. Future work suggested by the authors includes integrating generative models (VAEs, GANs) for more realistic modifications and amortized generation techniques to reduce computational overhead.

Technical Summary: Multi-SpaCE

More like this