Counterfactual explanations are essential for inter-preting predictions made by time series classifiers, yet existing methods often disrupt inherent temporal dependencies, resulting in unrealistic explanations. We propose CACTUS (Cross-Aligned Counterfactuals for Time Series Classification), a novel model-agnostic framework designed to generate interpretable explanations for black-box models. CACTUS leverages Dynamic Time Warping (DTW) Barycenter Averaging and introduces a cross-correlation loss to enforce temporal consistency and alignment. We introduce two variants of CACTUS: one that emphasizes minimal perturbations while maintaining classification validity, and another that ensures gradual, realistic modifications in the time series. Comprehensive experiments on diverse benchmark datasets demonstrate that CACTUS outperforms state-of-the-art optimization-based methods by generating more reliable and interpretable counterfactual explanations, establishing it as a robust tool for post-hoc time series models analysis and explanations.

Understanding the predictions of time series classifiers is essential in high-stakes domains such as health-care, finance, and security. Counterfactual explanations, which identify minimal perturbations to input sequences that alter model predictions, provide intuitive and actionable insights. However, existing methods often struggle to generate realistic or diverse counterfactuals, especially for sequential data with strong temporal dependencies. In this work, we propose DiverseLCF, a post-hoc framework for generating diverse and plausible counterfactual explanations for time series classifiers. By optimizing in the latent space of a pretrained variational autoencoder (VAE), DiverseLCF ensures that generated counterfactuals remain temporally coherent and aligned with the data manifold. We formulate the generation process as a multi-objective optimization problem that balances validity, proximity, and diversity. Extensive experiments show that DiverseLCF consistently generates valid, plausible, and diverse coun-terfactuals with minimal proximity trade-offs, providing a robust and generalizable tool for interpreting time series models.

EXplainable Artificial Intelligence (XAI) methods are essential for understanding complex machine learning models. This paper presents Adaptive Counterfactual Explanations for Time Series (ACTS), a model-agnostic algorithm that improves the interpretability, sparsity, and contiguity of counterfactuals. ACTS leverages Barycenter Averaging (DBA) with Dynamic Time Warping (DTW) and a regularized loss function combining prediction, DBA, L1 regularization, sparsity, and smoothness terms. Empirical results on real-world datasets show ACTS generates high-quality explanations with fewer outliers and better interpretability than current methods, advancing XAI in time series analysis.

Artificial intelligence (AI) and algorithmic decision-making are profoundly shaping various aspects of society, with applications in healthcare, business, education, etc. As these systems become more integral to high-stakes decisions, concerns about their transparency and interpretability are growing. To address these concerns, explainable AI (XAI) methods have been developed, with counterfactual explanations emerging as a powerful tool. Counterfactuals help users understand AI decisions by demonstrating how small changes in input could alter the outcome, providing a clear and intuitive way to interpret AI behavior. Despite their potential, generating valid, interpretable, and efficient counterfactual explanations is particularly challenging in time series domains, where data points are interdependent. In this paper, we introduce a novel approach to counterfactual explanations guided by ShapeDTW Barycenter Averaging (ShapeDBA). By integrating ShapeDBA into the counterfactual generation process, we ensure that the produced explanations are not only valid and interpretable but also efficient to generate. Our approach provides counterfactuals that align closely with human intuition while maintaining the computational efficiency required for practical deployment. This work represents a significant step forward in the development of interpretable AI systems, particularly in the complex domain of time series analysis.

Abstract The opacity inherent in machine learning models presents a significant hindrance to their widespread incorporation into decision-making processes. To address this challenge and foster trust among stakeholders while ensuring decision fairness, the data mining community has been actively advancing the explainable artificial intelligence paradigm. This paper contributes to the evolving field by focusing on counterfactual generation for time series classification models, a domain where research is relatively scarce. We develop, a post-hoc, model agnostic counterfactual explanation algorithm that leverages the Matrix Profile to map time series discords to their nearest neighbors in a target sequence and use this mapping to generate new counterfactual instances. To our knowledge, this is the first effort towards the use of time series discords for counterfactual explanations. We evaluate our algorithm on the University of California Riverside and University of East Anglia archives and compare it to three state-of-the-art univariate and multivariate methods.

Data augmentation plays an important part in the current success of machine learning and deep learning models. In particular, state-of-the-art architectures in the image recognition field include data augmentation modules as an integral part. However, there is still room for progress in the time series domain. In this work, we introduce OptimAug, a novel method for time series data augmentation. We deviate from the current state-of-the-art comprised of random transformations, pattern mixing, generative models, and decomposition methods, to develop the first multiloss-based optimization method. We evaluate our method with its two variants on datasets from the University of California Riverside (UCR) archive and compare it to multiple baseline algorithms from the literature.

As the demand for interpretable machine learning approaches increases, there is an increasing need for human involvement to provide diverse explanations for model decisions. This is crucial for enhancing trust and transparency in AI-based systems, leading to the emergence of the Explainable Artificial Intelligence (XAI) field. In this paper, we design a novel counterfactual explanation model, CELS, which learns a saliency map for the interest of an instance and generates a counterfactual explanation guided by the learned saliency map. CELS adopts a gradient-based approach composed of three interdependent modules that combine to generate sparse counterfactual explanations that are easily understood by end users. To the best of our knowledge, this is the first attempt to guide the perturbation to generate a counterfactual explanation via a learned saliency map. To validate our model, we conducted experiments using five popular real-world time-series datasets obtained from the UCR repository. The experimental results demonstrate the superiority of our model in achieving higher sparsity, proximity, and interpretability of counterfactual explanations when compared to other state-of-the-art baselines.

The rapid upsurge of numerical sources of information and the growth of storage capacities in recent years has resulted in the collection of massive time series datasets. Inspired by association rule mining and other rule discovery algorithms, several approaches have been proposed in the literature to discover temporal association rules from time series data. These methods place interpretability at the top of their priorities and aim to provide domain experts with relevant and qualitative rules. In this paper, we aim to fill the gap between temporal association rule mining and time series classification t asks to increase the interpretability of current classification methods. We propose rule transform (RT), a novel algorithm for multivariate time series classification ( MTSC) t hat generates discriminative temporal rules for the sake of classification. RT generates a new feature space that represents the support of the mined temporal rules which can easily be qualitatively interpreted by domain experts. The algorithm uses Allen’s Interval Algebra to extract the most prominent temporal rules from a given dataset. To our knowledge, this is the first effort to use shapelets as a unit for temporal rule mining studies for the purpose of classification. We evaluate our algorithm on the UEA archive of multivariate time series. Results show that RT produces accuracies superior to state-of-the-art time series classification algorithms with the additional advantage of interpretability.

EXplainable Artificial Intelligence ( XAI) methods have gained much momentum lately given their ability to shed the light on the decision function of opaque machine learning models. There are two dominating XAI paradigms: feature attribution and counterfactual explanation methods. While the first family of methods explains why the model made a decision, counterfactual methods aim at answering what-if the input is slightly different and results in another classification decision. Most of the research efforts have focused on answering the why question for time series data modality. In this paper, we aim at answering the what-if question by finding a good balance between a set of desirable counterfactual explanation properties. We propose Shapelet-guided Counterfactual Explanation (SG-CF), a novel optimization-based model that generates interpretable, intuitive post-hoc counterfactual explanations of time series classification models that balance validity, proximity, sparsity, and contiguity. Our experimental results on nine real-world time-series datasets show that our proposed method can generate counterfactual explanations that balance all the desirable counterfactual properties in comparison with other competing baselines.