🤖 AI

C-SHAP for time series: An approach to high-level temporal explanations

This paper introduces C-SHAP, a concept-based explainable AI framework that leverages SHAP to provide high-level temporal pattern explanations for time series models, addressing the limitations of existing point- or subsequence-based methods in high-stakes domains like healthcare and industry.

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

Published 2026-04-24

📖 5 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 have a very smart, but secretive, robot assistant. This robot looks at a long, wiggly line of data (like a heartbeat monitor or a machine's vibration sensor) and makes a decision: "This person is walking," or "This engine will break in 10 days."

The problem? The robot is a "Black Box." It gives you the answer, but it won't tell you why. If you ask, "Why did you say the engine is failing?" it just points to a specific dot on the line and says, "Because of this point right here."

But that's not very helpful. A single dot doesn't tell the whole story. The real reason might be that the whole line is slowly going up (a trend), or that the size of the wiggles is getting bigger (a scale issue).

This paper introduces a new tool called C-SHAP to fix this. Think of C-SHAP as a translator that turns the robot's secret code into a story humans can understand.

The Core Idea: From "Pixels" to "Pictures"

Most old methods of explaining AI are like looking at a photo through a microscope. They zoom in on individual pixels (or data points) and say, "This pixel made the robot think it's a cat." But humans don't see photos as pixels; we see patterns. We see a "tail," a "whisker," or a "furry texture."

C-SHAP changes the game. Instead of looking at individual dots, it breaks the data line down into Concepts (high-level patterns) like:

Trend: Is the line going up or down overall?
Bias: Is the line sitting high or low on the graph?
Scale: Are the waves big and wild, or small and calm?
Frequency: Is the line wiggling fast or slow?

It then asks the robot: "If I remove the 'Trend' from this data, do you still make the same decision?" If the robot's answer changes, C-SHAP knows that the Trend was the most important clue.

Two Ways to Ask the Robot

The paper suggests two ways to use this translator, depending on how the robot was built:

1. The "Post-Hoc" Approach (The Detective)

How it works: The robot was already built and trained on raw data. We don't touch the robot. Instead, we act like detectives. We take the robot's input, mathematically strip away the "Trend," then strip away the "Scale," and see how the robot's answer changes.
Analogy: Imagine a chef who already cooked a stew. You can't change the recipe now, but you can taste the stew, then taste it again after removing the salt, then again after removing the pepper. You realize, "Ah! The salt was the main flavor!" This is great because you can use it on any existing robot.

2. The "Concept-Informed" Approach (The Teacher)

How it works: Before we even train the robot, we teach it using the concepts. Instead of feeding it raw squiggly lines, we feed it separate channels: one channel for "Trend," one for "Scale," one for "Noise."
Analogy: Instead of letting the chef guess the ingredients, we give them bowls labeled "Salt," "Pepper," and "Sugar" and say, "Make the stew using these." The chef learns to rely on these specific ingredients from the start. This might make the chef (the AI) even better at cooking, and it's easier to explain because we know exactly what ingredients it used.

Real-World Examples

The authors tested this on two very different jobs:

1. Human Activity Recognition (The "Step Counter")

The Job: A robot watches a sensor on a person's knee to guess if they are standing, walking, sitting, or lying down.
The Old Way: The robot points to a specific second where the sensor moved and says, "Because of this movement."
The C-SHAP Way: The robot says, "I know they are walking because the Scale (how big the leg swings) is huge, and the Bias (the average angle of the knee) is tilted forward." This makes perfect sense to a human doctor or trainer.

2. Predictive Maintenance (The "Engine Doctor")

The Job: A robot watches an airplane engine to predict when it will break.
The Old Way: The robot points to a tiny spike in pressure and says, "Because of this spike."
The C-SHAP Way: The robot says, "I know the engine will fail soon because the Trend of the pressure is slowly going up over time." This tells the mechanic exactly what to look for: a slow, steady rise, not just a random spike.

Why This Matters

Trust: When a doctor or an engineer understands why an AI made a decision, they trust it more.
Simplicity: It stops the AI from giving confusing answers like "Point #452 is important." Instead, it gives human answers like "The overall trend is the problem."
Flexibility: You can use this on almost any time-based data, from stock markets to weather patterns.

The Bottom Line

C-SHAP is like a translator that stops AI from speaking in "data points" and starts speaking in "stories." It takes complex, wiggly lines of data and breaks them down into the big-picture concepts (Trend, Scale, Frequency) that humans naturally use to understand the world. This makes AI less of a mysterious black box and more of a helpful partner you can actually talk to.

1. Problem Statement

In high-stakes domains like healthcare and industry, AI models often operate as "black boxes," lacking transparency. While Explainable AI (XAI) techniques exist, most methods for time series data are limited to point-based (individual time steps) or subsequence-based (short windows) explanations.

Limitations: These local approaches fail to capture high-level global patterns (e.g., overall trends, variance changes, or frequency shifts) that often drive model decisions.
Consequence: Point-based attributions can be scattered and difficult for humans to interpret, failing to align with domain expert reasoning which often relies on abstract concepts rather than raw data points.
Gap: There is a lack of standardized, concept-based XAI frameworks for time series that can attribute importance to high-level features (concepts) rather than raw inputs.

2. Methodology: C-SHAP

The authors propose C-SHAP, a concept-based extension of the SHAP (SHapley Additive exPlanation) framework tailored for time series.

Core Concept

Instead of attributing importance to individual data points ( $x_i$ ), C-SHAP attributes importance to concepts ( $c_i$ ), defined as high-level patterns extracted from the time series (e.g., Trend, Frequency, Variance).

Two Approaches

The paper formalizes two distinct workflows for C-SHAP:

Fully Post-Hoc (Concept-Agnostic):
- The model is trained on raw time series data (black box).
- Concepts are extracted after training via a decomposition function $d(\cdot)$ .
- To calculate SHAP values, concepts are masked (replaced with background data) and then re-aggregated into a synthetic time series to feed back into the model.
- Requirement: The decomposition function must be invertible (i.e., the original signal can be perfectly reconstructed from its components).
Partially Ante-Hoc (Concept-Informed):
- The model is trained directly on the extracted concepts as separate input channels.
- Masking is simpler: specific concept channels are replaced with background concept values without needing re-aggregation.
- Benefit: May improve model performance by explicitly guiding the model to learn relevant high-level features.

Concept Construction (Decomposition)

To generate concepts, the authors utilize Time Series Decomposition. They propose and test two methods:

Discrete Wavelet Transform (DWT): A standard mathematical approach splitting signals into Approximation (low frequency) and Detail (high frequency) components.
Custom Human-Centric Decomposition: A hybrid (additive and multiplicative) decomposition designed for interpretability. It extracts specific concepts:
- Trend: Linear slope.
- Bias: Average signal level.
- Scale: Maximum amplitude.
- Variance: Changes in amplitude over time.
- Low/High Frequency: Slow vs. rapid changes.

Masking Mechanism

SHAP requires calculating the model output when features are "missing." In C-SHAP:

Concepts are masked by replacing them with samples drawn from the training distribution (background data).
For the Concept-Agnostic approach, the masked components are recombined using the inverse of the decomposition function to create a valid input signal for the black-box model.

3. Key Contributions

Formalization of C-SHAP: Defined two rigorous approaches (post-hoc and ante-hoc) for applying SHAP to high-level time series concepts.
Invertibility Condition: Established a mathematical condition (Condition 2) requiring concept construction to be invertible for post-hoc applications to ensure valid SHAP values.
Custom Decomposition Algorithm: Designed a novel, human-interpretable decomposition method that isolates specific physical properties (Trend, Scale, Variance) rather than just mathematical frequencies.
Empirical Validation: Demonstrated the framework across two distinct domains:
- Human Activity Recognition (HAR): Classifying locomotion (Stand, Walk, Sit, Lie).
- Predictive Maintenance: Estimating Remaining Useful Life (RUL) of aircraft turbofan engines.

4. Results

The authors evaluated C-SHAP against standard point-based KernelSHAP using GRU-based models.

Interpretability:
- HAR (OPPORTUNITY Dataset): C-SHAP revealed that models relied heavily on "Bias" (average knee angle) and "Scale" (amplitude of movement) to distinguish activities. Point-based methods failed to clearly identify these global patterns, showing scattered importance across time steps.
- Predictive Maintenance (Turbofan Dataset): C-SHAP identified "Trend" (linear pressure change) as the most critical factor for predicting engine failure. KernelSHAP could not isolate this global trend due to the distributed nature of the attribution.
Performance:
- Concept-informed models (trained on concepts) showed comparable or slightly improved performance over concept-agnostic models in some cases, suggesting that feeding high-level features directly to the model can be beneficial.
- The custom decomposition provided more intuitive explanations for lay users compared to the mathematically rigorous but less intuitive DWT.

5. Significance

Bridging the Gap: C-SHAP moves XAI beyond local, pixel/point-level explanations to semantic, high-level reasoning, aligning better with how human experts analyze time series data.
Flexibility: The framework is model-agnostic and can be applied to any time series model, provided the concept extraction is invertible (for post-hoc use).
Domain Applicability: By successfully applying the method to both classification (HAR) and regression (Predictive Maintenance), the paper proves the generalizability of concept-based explanations across different time series tasks.
Future Direction: The work highlights the need for user studies to validate if these high-level explanations actually increase trust and usability for end-users and domain experts.

In summary, C-SHAP offers a robust framework for explaining time series AI by shifting the focus from "which time point mattered" to "which pattern (trend, frequency, variance) mattered," thereby enhancing the transparency and trustworthiness of AI in critical applications.