CONCEPTCF: CONCEPT-BASED COUNTERFACTUALS FOR THE EXPLAINABILITY OF TIME SERIES

As artificial intelligence models are increasingly deployed in high-stakes domains like healthcare and industrial maintenance, understanding their decision-making process is critical. While counterfactual explanations—which identify the minimal changes required to alter a model's prediction—are effective for this purpose, existing methods for time series often operate on raw data points or subsequences. This approach frequently produces non-interpretable results or violates the temporal dependencies of the data. The authors seek to address this by developing a method that generates counterfactuals based on human-interpretable concepts.

The authors introduce ConceptCF, a framework that shifts the focus of counterfactual generation from raw input features to high-level concepts. The process involves three stages: first, decomposing the time series into interpretable components (such as scale, trend, or frequency bands) using techniques like Fourier or Wavelet transforms; second, using a genetic algorithm to optimize these concepts to find a counterfactual that changes the model's prediction; and third, reconstructing the time series from the modified concepts. By optimizing at the concept level, the method ensures that the resulting explanations are intuitive, such as stating that a prediction would change if the 'scale' of a movement were increased.

ConceptCF was evaluated against five state-of-the-art counterfactual generation methods across several datasets, including MotionSenseHAR and various UCR archive datasets. The results demonstrate that ConceptCF consistently achieves top-tier performance across key metrics: validity (ensuring the prediction changes), confidence, proximity (resembling the original sample), sparsity (minimizing the number of changes), and plausibility (staying within the data distribution). The authors show that by using a genetic algorithm to perturb concepts, they can provide meaningful, contrastive explanations that align with human reasoning while maintaining the structural integrity of the time series.