IMFACT: Counterfactual Explanations for Time Series via IMF Substitution
A recent preprint on arXiv (2608.04777) presents IMFACT, a versatile framework designed for creating credible counterfactual explanations tailored for time series classifiers. This approach functions within the decomposition realm of Empirical Mode Decomposition (EMD), where input signals are divided into Intrinsic Mode Functions (IMFs). It systematically replaces chosen IMFs with those from a Nearest Unlike Neighbour (NUN) until the classifier transitions to the desired class. The research assesses six strategies for IMF selection and a multi-NUN cycling method across two UCR benchmarks (FaultDetectionA and FruitFlies). Notably, the variance-based strategy utilizing three NUNs surpasses two established baseline methods in terms of reliability and plausibility. By working within the decomposition space, IMFACT maintains temporal integrity, producing more realistic counterfactuals. This preprint is classified as a cross-type announcement and can be found on arXiv.
Key facts
- IMFACT is a model-agnostic framework for counterfactual explanations in time series classification.
- It operates in the decomposition space of Empirical Mode Decomposition (EMD).
- Input signals are split into Intrinsic Mode Functions (IMFs).
- Selected IMFs are substituted with those from a Nearest Unlike Neighbour (NUN) until the classifier flips to the target class.
- Six IMF-selection strategies and a multi-NUN cycling extension were evaluated.
- Benchmarks used: UCR FaultDetectionA and FruitFlies datasets.
- The variance-based strategy with three NUNs outperformed two baseline techniques on reliability and plausibility.
- The paper is available on arXiv with ID 2608.04777.
Entities
Institutions
- arXiv