Perturbation-based time series interpretation suffers under two challenges: firstly, the long and multivariate time series may lead to various incoherent salient spots on the saliency map, and secondly, common perturbation techniques often return unrealistic sequences. In this paper, we propose Cohesive Explanation for Time Series (CETS). This time series interpretation approach provides cohesive (a notion of concentrated salient features at adjacent timestamps) feature attribution using realistic prototype-based perturbations. CETS ensures a cohesive interpretation by employing both global (temporal) and local (spatial) perturbations of time series. These perturbations confine the interpretation to a concise temporal event within a specific subspace. We perform extensive experiments on real-world benchmark datasets to demonstrate the efficacy of our interpretations. Specifically, we visually illustrate how cohesive attributions contribute to enhancing the interpretability of intricate time series data. Our empirical results show that CETS achieves interpretation quality comparable to state-of-the-art approaches while providing cohesive and easy-to-understand explanations.

Detecting temporal abnormal patterns over streaming data is challenging due to volatile data properties and the lack of real-time labels. The abnormal patterns are usually hidden in the temporal context, which cannot be detected by evaluating single points. Furthermore, the normal state evolves over time due to concept drifts. A single model does not fit all data over time. Autoencoders have recently been applied for unsupervised anomaly detection . However, they are trained on a single normal state and usually become invalid after distributional drifts in the data stream. This paper uses an Autoencoder-based approach STAD for anomaly detection under concept drifts. In particular, we propose a state-transition-aware model to map different data distributions in each period of the data stream into states, thereby addressing the model adaptation problem in an interpretable way. In addition, we analyzed statistical tests to detect the drift by examining the sensitivity and powers. Furthermore, we present considerable ways to estimate the probability density function for comparing the distributional similarity for state transitions. Our experiments evaluate the proposed method on synthetic and real-world datasets. While delivering comparable anomaly detection performance as the state-of-the-art approaches, STAD works more efficiently and provides extra interpretability . We also provide insightful analysis of optimal hyperparameters for efficient model training and adaptation.

For volatile multivariate time series, variations in the distributions of the input dimension and the correlation structure present an open challenge. The different distributions before and after a change-point hinder the performance of most of the predictive methods, mostly requiring re-training of the models. The detection of such change points represents a severe problem, as volatile data labeling is often either expensive or delayed in streaming data; Moreover, classical concept drift detectors usually struggle with detecting changes in correlations of multivariate time series’ input variables. We focus on unsupervised change detection, tracking correlation changes in the input variables without class labels. By introducing slidSHAPs, we propose a fully unsupervised change detector for multivariate time series with categorical value domains; our tool detects correlation-based changes through a representation of the correlation structure of the input data. The slidSHAPs series underlines distributional changes even in a few univariate input variables, thus, being more sensitive to changes than any prior change point detection method. In contrast to the well-known application of Shapley values for interpretable machine learning, we use this foundational game-theoretic concept to extrapolate information on the correlation structure of data streams and achieve higher sensitivity towards multiple changes in the empirical evaluation of synthetic and real-world data.

Detecting anomalies in time series data is challenging due to their complex and volatile temporal features. Some anomalies only show deviating patterns to their local context instead of the overall distribution. Additionally, the biased sample distribution between normal and abnormal classes hinders the efficient usage of the available data labels. Self-supervised approaches are practically efficient for anomaly detection, in which only normal data is used during the training. However, they often fail to detect contextual anomalies in high-dimensional time series data, while the representation learning of such complex data patterns is sub-optimal. This paper introduces ContrastAD, a novel self-supervised framework for time series anomaly detection. Specifically, we employ the contrastive learning process with anomaly-induced temporal transformations. Targeting the point and contextual anomalies appearing in time series data, we develop corresponding transformations to enforce the model to learn discrepant representations for normal and abnormal data in the latent space. With extensive experiments, we show that our approach outperforms baseline anomaly detectors on various benchmark datasets. Our empirical results indicate that ContrastAD improves anomaly detection performance on noisy and high-dimensional time series datasets, even without common repeating patterns.