🤖 machine learning

ConceptCF: Concept-based Counterfactuals for the Explainability of Time Series

This paper introduces ConceptCF, a novel method for generating interpretable counterfactual explanations in time series by modifying human-understandable concepts derived from time series decomposition, which outperforms existing approaches across key evaluation metrics.

Original authors: Annemarie Jutte, Faizan Ahmed, Jeroen Linssen, Maurice van Keulen

Published 2026-07-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Annemarie Jutte, Faizan Ahmed, Jeroen Linssen, Maurice van Keulen

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 trying to teach a robot how to understand the world. You give it a camera and a brain, and it starts making predictions: "That's a cat," or "That machine is about to break." But sometimes, the robot gets it wrong, or worse, it gets it right for the wrong reasons. Maybe it thinks a picture of a wolf is real because there's snow in the background, not because it actually sees the wolf. This is a big problem in the world of Artificial Intelligence (AI), especially when the stakes are high, like in hospitals or factories. We need to know why the robot made a decision, not just what it decided.

To fix this, scientists use something called "Explainable AI." One of the coolest tools in this toolbox is the "counterfactual." Think of a counterfactual as a "what if" story. Instead of just saying, "I predicted you were walking," a counterfactual says, "I predicted you were walking, but if you had slowed down your steps, I would have predicted you were sitting." It's like a recipe for changing the future. However, when dealing with time series—data that changes over time, like a heartbeat monitor or a stock chart—making these "what if" stories is tricky. If you just tweak a single point in the data, it's like changing one word in a sentence and hoping the whole story still makes sense. Often, it doesn't. The story becomes gibberish, and the robot's explanation becomes useless.

This is where a new method called ConceptCF comes in. The researchers behind this paper, Annemarie Jutte and her team, realized that instead of fiddling with individual points in a time series, we should change the "big ideas" or "concepts" that make up the data. Imagine a song. You could try to change the song by tweaking every single note one by one, which is messy and confusing. Or, you could change the "volume" (how loud it is) or the "tempo" (how fast it is). These are the concepts. ConceptCF is a clever system that takes a time series, breaks it down into these understandable concepts (like how big the waves are or how often they repeat), and then asks: "What if we changed the volume or the tempo?" It uses a digital evolution process, like a video game where the best solutions survive and get better, to find the smallest, most logical changes needed to flip a prediction. The result is an explanation that a human can actually understand, like saying, "The robot thought you were walking because your steps were too big; if you took smaller steps, it would know you were sitting."

The paper introduces ConceptCF as a way to generate these "what if" explanations specifically for time series data. The core idea is to stop looking at the raw data points (the individual moments in time) and start looking at concepts—high-level patterns that humans can easily grasp, such as the overall "scale" (amplitude) of a signal or its "frequency" (how often it repeats). The authors argue that existing methods often mess up the meaning of the data by changing points in isolation, which can create weird, impossible scenarios. ConceptCF avoids this by only modifying the underlying concepts.

To do this, the team uses three different ways to break down the time series into concepts:

Fourier-based: This looks at the signal as a mix of different frequencies, like separating a chord into individual notes.
Wavelet-based: This breaks the signal into chunks of different sizes, capturing both big trends and small details.
Human-centered: This is designed specifically for people who aren't math experts, breaking the signal into things like "trend" (is it going up or down?), "bias" (is it shifted up or down?), and "scale" (how big is it?).

Once the data is broken into these concepts, the method uses a genetic algorithm to find the perfect counterfactual. Think of this like a digital breeding program. The computer starts with a "population" of possible changes. It tests them to see which ones successfully change the AI's prediction (validity) while staying as close to the original data as possible (proximity) and changing as few things as possible (sparsity). The "fittest" solutions survive and mix together to create even better solutions in the next generation. The goal is to find the minimal change in concepts that flips the prediction. For example, if a model predicts "Walk," ConceptCF might find that increasing the "wavelength" and "scale" of the movement would make the model predict "Sit."

The researchers tested ConceptCF against five other top-tier methods using real-world datasets, including motion data from people walking and heart rate data. They measured how well the new method worked based on five criteria:

Validity: Did it actually change the prediction?
Confidence: Was the model sure about the new prediction?
Proximity: Was the new scenario close to the original?
Sparsity: Did it change only a few things, or did it mess with everything?
Plausibility: Did the new scenario look like something that could actually happen in the real world?

The results showed that ConceptCF performed very competitively. In fact, it often came out on top, especially for sparsity. This means it could explain the AI's decision by changing very few concepts, making the explanation much cleaner and easier for a human to understand. While other methods sometimes struggled to keep the data looking realistic or ended up changing too many points, ConceptCF consistently found the "sweet spot" of making the smallest, most logical change to flip the outcome. The authors found that the specific way they broke down the concepts (Fourier vs. Wavelet vs. Human-centered) didn't drastically change the results, suggesting that the method is flexible and can be adapted to different needs.

The paper concludes that ConceptCF is a powerful new tool for understanding time series AI. By focusing on human-interpretable concepts rather than raw data points, it bridges the gap between complex machine learning models and human reasoning. The authors suggest that while their method is mathematically sound and performs well on standard tests, the ultimate test is whether humans actually understand the explanations better. They propose that future research should involve real people to see if these concept-based explanations truly make AI more transparent and trustworthy. For now, ConceptCF stands as a promising step toward making AI decisions less like a black box and more like a conversation we can all understand.

Technical Summary: ConceptCF for Time Series Explainability

Problem Definition

In high-stakes domains such as healthcare and predictive maintenance, Artificial Intelligence (AI) models often rely on spurious correlations rather than causal relationships, necessitating Explainable AI (XAI). While counterfactual explanations—identifying minimal modifications to change a model's prediction—are a popular XAI technique, existing methods for time series data face significant challenges. Standard approaches typically modify individual data points or subsequences in isolation. This point-wise optimization risks violating temporal dependencies, generating out-of-distribution samples, and altering the semantic meaning of the sequence in ways that are difficult for humans to interpret. Furthermore, global characteristics of time series, such as scale or periodicity, are distributed across the signal and cannot be adequately captured by modifying a few points.

The paper addresses the need for counterfactual generation that operates on human-interpretable concepts rather than raw data points, ensuring that the resulting explanations clarify which high-level data properties drive model decisions.

Methodology: ConceptCF

The proposed method, ConceptCF, generates counterfactuals by perturbing high-level concepts derived from time series decomposition, rather than the raw signal itself. The methodology proceeds in three stages:

Concept Construction: A time series $x$ is decomposed into a set of concepts $S_x = \{c_1, ..., c_m\}$ using an invertible function $d(\cdot)$ . The paper evaluates three decomposition strategies:
- Discrete Fourier Transform (DFT): Decomposes the signal into frequency bands.
- Discrete Wavelet Transform (DWT): Decomposes the signal into approximation and detail components.
- Human-Centered Decomposition: Extracts semantic components including linear trend, bias (mean), scale (amplitude), low-frequency components, sliding moving variance, and high-frequency components.
- Constraint: The decomposition must be invertible ( $x = h(S_x)$ ) to reconstruct a meaningful signal from perturbed concepts.
Optimization Problem: The generation of a counterfactual is formulated as a multi-objective optimization problem. The goal is to find a perturbed concept set $S_x^{CF}$ that minimizes a weighted sum of four objectives:
- Confidence: Maximizing the model's probability for the target class.
- Proximity: Minimizing the L1-norm distance between the original and counterfactual samples (normalized by Mean Absolute Deviation).
- Sparsity: Minimizing the fraction of concepts changed.
- Plausibility: Minimizing the increase in outlier score (IOS) relative to a trained autoencoder, ensuring the counterfactual remains within the data distribution.
- Validity Constraint: The counterfactual must result in a different predicted class ( $y_{x^{CF}} \neq y_x$ ).
Genetic Algorithm (GA): The optimization is solved using a genetic algorithm.
- Initialization: The population is seeded with the Nearest Unlike Neighbour (NUN) to ensure validity.
- Mutation: Scalar concepts are perturbed with Gaussian noise (clipped to training ranges), while sequential concepts are replaced by nearest neighbours from the training set to preserve semantic meaning.
- Elitism: The fittest individuals are preserved across generations to guarantee that the validity constraint is maintained throughout the optimization process.

Key Contributions

The paper presents the following contributions to the state of the art:

Formalization: A formal definition of the counterfactual problem for time series within the context of concept-based XAI.
Algorithm: The development of ConceptCF, a method solving the concept-based counterfactual problem using a genetic algorithm.
Demonstration: A demonstration of how concepts (e.g., scale, frequency, trend) can be used to generate meaningful counterfactuals, allowing explanations such as "the prediction would change if the scale of the movement increased."
Validation: An empirical validation showing that ConceptCF performs competitively against existing methods across standard metrics.

Experimental Results

The authors evaluated ConceptCF against five state-of-the-art baselines (Wachter, Native Guide, TSEvo, Glacier, and Sub-SpaCE) using five datasets (MS-HAR, CBF, Coffee, ECG200, GunPoint).

Validity: ConceptCF achieved 100% validity across all datasets, guaranteed by the inclusion of the NUN in the initial population and the preservation of valid solutions via elitism.
Sparsity: ConceptCF achieved the highest rank (best performance) for sparsity across all datasets. This indicates that ConceptCF requires fewer changes (modifying fewer high-level concepts) to alter a prediction compared to point-based methods.
Proximity: ConceptCF also achieved the top rank for proximity, suggesting the generated counterfactuals remain close to the original samples.
Confidence and Plausibility: ConceptCF ranked second in these metrics. The authors note that Glacier performed slightly better in plausibility, likely due to its direct generation in the latent space of an autoencoder.
Decomposition Comparison: The three decomposition approaches (Human-centered, DFT, DWT) performed comparably, with the Human-centered approach showing slightly better sparsity results.

Significance and Claims

The paper claims that ConceptCF offers a significant advancement in time series explainability by shifting the focus from opaque point-wise modifications to interpretable, high-level concepts.

Interpretability: By operating on concepts like "scale" or "frequency," the method produces explanations that align with human reasoning (e.g., "increase the scale to change the prediction"), addressing the "black box" nature of time series models.
Competitive Performance: The authors assert that ConceptCF is not only interpretable but also technically competitive, matching or exceeding the performance of existing methods in validity, confidence, proximity, and sparsity.
Modesty: The authors acknowledge that while the metrics demonstrate technical competence, they do not fully capture user comprehension. Consequently, they suggest that user evaluation is necessary for future research to validate the practical interpretability of the generated explanations. They also note that the choice of decomposition should be aligned with the specific domain and end-user needs.