Backdoor attacks are a significant threat to the performance and integrity of pre-trained language models. Although such models are routinely fine-tuned for downstream NLP tasks, recent work shows they remain vulnerable to backdoor attacks that survive vanilla fine-tuning. These attacks are difficult to defend because end users typically lack knowledge of the attack triggers. Such attacks consist of stealthy malicious triggers introduced through subtle syntactic or stylistic manipulations, which can bypass traditional detection and remain in the model, making post-hoc purification essential. In this study, we explore whether attention-head pruning can mitigate these threats without any knowledge of the trigger or access to a clean reference model. To this end, we design and implement six pruning-based strategies: (i) gradient-based pruning, (ii) layer-wise variance pruning, (iii) gradient-based pruning with structured L1/L2 sparsification, (iv) randomized ensemble pruning, (v) reinforcement-learning-guided pruning, and (vi) Bayesian uncertainty pruning. Each method iteratively removes the least informative heads while monitoring validation accuracy to avoid over-pruning. Experimental evaluation shows that gradient-based pruning performs best while defending the syntactic triggers, whereas reinforcement learning and Bayesian pruning better withstand stylistic attacks.

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.

Generating high-quality synthetic time series is a fundamental yet challenging task across domains such as forecasting and anomaly detection, where real data can be scarce, noisy, or costly to collect. Unlike static data generation, synthesizing time series requires modeling both the marginal distribution of observations and the conditional temporal dependencies that govern sequential dynamics. We propose TIMED, a unified generative framework that integrates a denoising diffusion probabilistic model (DDPM) to capture global structure via a forward-reverse diffusion process, a supervisor network trained with teacher forcing to learn autoregressive dependencies through next-step prediction, and a Wasserstein critic that provides adversarial feedback to ensure temporal smoothness and fidelity. To further align the real and synthetic distributions in feature space, TIMED incorporates a Maximum Mean Discrepancy (MMD) loss, promoting both diversity and sample quality. All components are built using masked attention architectures optimized for sequence modeling and are trained jointly to effectively capture both unconditional and conditional aspects of time series data. Experimental results across diverse multivariate time series benchmarks demonstrate that TIMED generates more realistic and temporally coherent sequences than state-of-the-art generative models.

Data augmentation can significantly enhance the performance of machine learning tasks by addressing data scarcity and improving generalization. However, generating time series data presents unique challenges. A model must not only learn a probability distribution that reflects the real data distribution but also capture the conditional distribution at each time step to preserve the inherent temporal dependencies. To address these challenges, we introduce AVATAR, a framework that combines Adversarial Autoencoders (AAE) with Autoregressive Learning to achieve both objectives. Specifically, our technique integrates the autoencoder with a supervisor and introduces a novel supervised loss to assist the decoder in learning the temporal dynamics of time series data. Additionally, we propose another innovative loss function, termed distribution loss, to guide the encoder in more efficiently aligning the aggregated posterior of the autoencoder’s latent representation with a prior Gaussian distribution. Furthermore, our framework employs a joint training mechanism to simultaneously train all networks using a combined loss, thereby fulfilling the dual objectives of time series generation. We evaluate our technique across a variety of time series datasets with diverse characteristics. Our experiments demonstrate significant improvements in both the quality and practical utility of the generated data, as assessed by various qualitative and quantitative metrics.

Current Generative Adversarial Network (GAN)-based approaches for time series generation face challenges such as suboptimal convergence, information loss in embedding spaces, and instability. To overcome these challenges, we introduce an advanced framework that integrates the advantages of an autoencoder-generated embedding space with the adversarial training dynamics of GANs. This method employs two discriminators: one to specifically guide the generator and another to refine both the autoencoder’s and generator’s output. Additionally, our framework incorporates a novel autoencoder-based loss function and supervision from a teacher-forcing supervisor network, which captures the stepwise conditional distributions of the data. The generator operates within the latent space, while the two discriminators work on latent and feature spaces separately, providing crucial feedback to both the generator and the autoencoder. By leveraging this dual-discriminator approach, we minimize information loss in the embedding space. Through joint training, our framework excels at generating high-fidelity time series data, consistently outperforming existing state-of-the-art benchmarks both qualitatively and quantitatively across a range of real and synthetic multivariate time series datasets.

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.

In heliophysics research, predicting solar flares is crucial due to their potential to substantially impact both space-based systems and Earth’s infrastructure. Magnetic field data from solar active regions, recorded by solar imaging observatories, are transformed into multivariate time series to enable solar flare prediction using temporal window-based analysis. In the realm of multivariate time series-driven solar flare prediction, addressing severe class imbalance with effective strategies for multivariate time series representation learning is key to developing robust predictive models. Traditional methods often struggle with overfitting to the majority class in prediction tasks where major solar flares are infrequent. This work presents EX-CON, a contrastive representation learning framework designed to enhance classification performance amidst such imbalances. EXCON operates through four stages: (1) obtaining core features from multivariate time series data; (2) selecting distinctive contrastive representations for each class to maximize inter-class separation; (3) training a temporal feature embedding module with a custom extreme reconstruction loss to minimize intra-class variation; and (4) applying a classifier to the learned embeddings for robust classification. The proposed method leverages contrastive learning principles to map similar instances closer in the feature space while distancing dissimilar ones, a strategy not extensively explored in solar flare prediction tasks. This approach not only addresses class imbalance but also offers a versatile solution applicable to both univariate and multivariate time series across binary and multiclass classification problems. Experimental results, including evaluations on the benchmark solar flare dataset and multiple time series archive datasets with binary and multiclass labels, demonstrate EXCON’s efficacy in enhancing classification performance and reducing overfitting.

Minority stress is the leading theoretical construct for understanding LGBTQ+ health disparities. As such, there is an urgent need to develop innovative policies and technologies to reduce minority stress. To spur technological innovation, we created the largest labeled datasets on minority stress using natural language from subreddits related to sexual and gender minority people. A team of mental health clinicians, LGBTQ+ health experts, and computer scientists developed two datasets: (1) the publicly available LGBTQ+ Minority Stress on Social Media (MiSSoM) dataset and (2) the advanced request-only version of the dataset, LGBTQ+ MiSSoM+. Both datasets have seven labels related to minority stress, including an overall composite label and six sublabels. LGBTQ+ MiSSoM (N = 27,709) includes both human- and machine-annotated la-bels and comes preprocessed with features (e.g., topic models, psycholinguistic attributes, sentiment, clinical keywords, word embeddings, n-grams, lexicons). LGBTQ+ MiSSoM+ includes all the characteristics of the open-access dataset, but also includes the original Reddit text and sentence-level labeling for a subset of posts (N = 5,772). Benchmark supervised machine learning analyses revealed that features of the LGBTQ+ MiSSoM datasets can predict overall minority stress quite well (F1 = 0.869). Benchmark performance metrics yielded in the prediction of the other labels, namely prejudiced events (F1 = 0.942), expected rejection (F1 = 0.964), internalized stigma (F1 = 0.952), identity concealment (F1 = 0.971), gender dysphoria (F1 = 0.947), and minority coping (F1 = 0.917), were excellent. Descriptive analyses, ethical considerations, limitations, and possible use cases are provided.

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.

LGBTQ+ minority stress is a pervasive form of anti-LGBTQ+ adverse events and psychological strain that drives health inequities among LGBTQ+ people. Minority stress is also linguistically sophisticated (e.g., composed of cultural idioms, psycholinguistic permutations, and lexical density). Because minority stress is a linguistically sophisticated social determinant of health disparities, it is challenging to detect using natural language processing (NLP). Using 5,789 human-annotated Reddit posts from the LGBTQ+ Minority Stress on Social Media (MiSSoM+) Dataset, we investigated and compared the performance of four neural networks and two traditional machine learning architectures in modeling minority stress at both the factor (i.e., separate components of minority stress) and composite level. A novel hybrid model combining Bidirectional Encoder Representations from Transformers and convolutional neural network (BERT-CNN) improved the prediction of composite minority stress (F1 = 0.84). Our experiments on separate factors of minority stress are the first to demonstrate that hybrid neural network models can detect semantically complex expressions of prejudiced events (F1 = 0.87), expected rejection (F1 = 0.92), internalized stigma (F1 = 0.91), identity concealment (F1 = 0.92), and minority coping (F1 = 0.84). We also substantially improved the prediction of gender dysphoria (F1 = 0.94)—a conceptually new candidate component of minority stress. Big data analytics may not be a panacea for the problem of minority stress, but our work joins a growing literature base to show that deep learning models are remarkable in detecting linguistically sophisticated social determinants of health disparities in big data, thus providing evidence in support of the potential benefit from the innovative use of such technology in eliminating group-specific health inequities.

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.

Solar physicists frequently use solar magnetic field parameters for analyzing and predicting solar events. Temporal observation of magnetic field parameters, i.e., multivariate time series (MVTS) representation facilitates finding relationships of magnetic field states to the occurrence of extreme solar events (e.g., solar flares). Feature selection of MVTS-represented solar magnetic field parameters (features) can select the most relevant parameters that give high prediction accuracy. In this paper, we propose a deep learning-based feature selection method, more specifically, an LSTM-based incremental feature selection method, as an end-to-end solution for feature selection in MVTS data. We performed LSTM-based feature selection for multivariate time series data in two steps. Firstly, each MVTS feature is evaluated individually by an LSTM-based univariate sequence classifier, and secondly, the top-performing features are combined to produce input for a downstream LSTM-based multivariate sequence classifier. We compared the proposed MVTS feature selection method with three other baseline feature selection methods on an MVTS-based solar flare prediction dataset and demonstrated that our method selects more discriminatory features compared to other methods.

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.

EXplainable Artificial Intelligence (XAI) methods are increasingly accepted as effective tools to trace complex machine learning models' decision-making processes. There are two underlying XAI paradigms: (1) traditional factual methods and (2) emerging counterfactual models. The first family of methods uses feature attribution techniques that alter the feature space and observe the impact on the decision function. Counterfactual models aim at providing the smallest possible change to the feature vector that can change the prediction outcome. In this paper, we propose TimeX, a new model-agnostic time series counterfactual explanation algorithm that provides sparse, interpretable, and contiguous explanations. We validate our model using real-world time series datasets and show that our approach can generate explanations with up to 20% fewer outliers in comparison with other state-of-the-art competing baselines.

Clustering is an important data mining task that consists of grouping data without a prior knowledge of classes. In the recent years, with the rise of big data and its various applications, unsupervised learning such as clustering algorithms have attracted massive interest in the data mining field. For the case of time series data, temporal changes are difficult to detect, which makes the clustering for temporal sequence data more complex than traditional vector data. In this paper, we propose a new model that generates graphs from time series data to preserve important relations between different data points. In particular, every time series data will be considered as a node, then edges will be added between nodes if the dynamic time warping distances of time series data achieve a specific threshold, then spectral clustering algorithm is applied to the generated graph. Our results shows that our new proposal time series graph representation outperforms state-of-the-art clustering algorithms.

Shapelets, also known as motifs, are time series sequences that have the property of discriminating between time series classes. Lately, shapelets studies have gained a lot of momentum due to their interpretable nature. As opposed to traditional time series classifiers, shapelet-based learners provide a visual representation of the pattern that triggers the classification decision. One of the most challenging issues of shapelet-based classifiers is the generation of a large number of shapelet outputs. To the best of our knowledge, this is the first effort that addresses the high numerosity problem of mined shapelets issue by mining the minimal set of discriminative shapelets for time series data. We propose a new shapelet mining learner, 1DCNN, that has the property of learning shapelets of different lengths using a black-box neural network model. 1DCNN optimizes the entire classification schema by learning the shapes of the representative patterns. Our proposed model uses network pruning to sparsify the network and keep only the most discriminative shapelets without compromising the classification accuracy. We validated our model using 59 real-world time series datasets from the UCR repository. Our experimental results show the effectiveness and efficiency of our approach in comparison with other competing baselines models. For fairness purposes, we did not compare 1DCNN with ensemble based approaches that encapsulates many learners. Our results show that the performance of our model is superior to all other baselines pertaining to the shapelet-based classifier category, with up to 95% less Floating Points Operations per Second (FLOPs) required by the network.

In recent years, node embedding algorithms, which learn low dimensional vector representations for nodes in a graph, have been one of the key research interests of the graph mining community. The existing algorithms either rely on computationally expensive eigendecomposition of the large matrices, or require tuning of the word embedding-based hyperparameters as a result of representing the graph as a node sequence similar to the sentences in a document. Moreover, the latent features produced by these algorithms are hard to interpret. In this paper, we present Tensor Decomposition-based Node Embedding (TDNE), a novel model for learning node representations for arbitrary types of graphs: undirected, directed, and/or weighted. Our model preserves the local and global structural properties of a graph by constructing a third-order tensor using the k-step transition probability matrices and decomposing the tensor through CANDECOMP/PARAFAC (CP) decomposition in order to produce an interpretable, low dimensional vector space for the nodes. Our experimental evaluation using two well-known social network datasets proves TDNE to be interpretable with respect to the understandability of the feature space, and precise with respect to the network reconstruction.

In recent years, node embedding algorithms, which learn low dimensional vector representations for nodes in a graph, have been one of the key research interests of the graph mining community. The existing algorithms either rely on computationally expensive eigendecomposition of the large matrices, or require tuning of the word embedding-based hyperparameters as a result of representing the graph as a node sequence similar to the sentences in a document. Moreover, the latent features produced by these algorithms are hard to interpret. In this paper, we present two novel tensor decomposition-based node embedding algorithms, that can learn node features from arbitrary types of graphs: undirected, directed, and/or weighted, without relying on eigendecomposition or word embedding-based hyperparameters. Both algorithms preserve the local and global structural properties of the graph by using k-step transition probability matrices to construct third-order multidimensional arrays or tensors and perform CANDECOMP/PARAFAC (CP) decomposition in order to produce an interpretable and low dimensional vector space for the nodes. Our experiments encompass different types of graphs (undirected/directed, unweighted/weighted, sparse/dense) of different domains such as social networking and neuroscience. Our experimental evaluation proves our models to be interpretable with respect to the understandability of the feature space, precise with respect to the network reconstruction and link prediction, and accurate with respect to node classification and graph classification.

Solar flare prediction is an important task because of their potential impacts on both space and terrestrial infrastructure. This prediction task can be modeled as a binary classification between flaring and non-flaring Active Regions. Previous works on flare prediction focused on representing flaring and non-flaring Active Region examples in vector space, where the feature space was found from the Active Region magnetic field parameters. We extract time series samples of these Active Region parameters and present a flare prediction method based on the k-NN classification of the univariate time series. We find that, for our classification task, using a statistical summarization on the time series of a single Active Region parameter, called total unsigned current helicity, outperforms the use of all Active Region parameters at a single instant of time. Additionally, we present a data model of the flaring/non-flaring Active Regions using multivariate time series.

Space Weather is of rising importance in scientific discipline that describes the way in which the Sun and space impact a myriad of activities down on Earth as well as the safety of the space crew members on board of the space stations. Consequently, it is imperative to better quantify the risk of future space weather events. Most of the flare prediction models in literature use physical parameters of the potentially flaring active regions during a limited interval to gain insights on whether a flare will happen or not. This limits our perception of how an event evolves for an extended duration across multiple parameters. In this paper we followed a data-driven approach to address the problem of flare prediction from a multivariate time series analysis perspective and attempt to cluster potential flaring active regions by applying Distance Density clustering on individual parameters and further organize the clustering results into a multivariate time series decision tree. We compared different data extraction priors and spans, and ranked the importance for different parameters through univariate clustering. To the best of our knowledge, this is the first attempt to predict solar flares using a tree structure.