Robust Evaluation of Attribution Methods for Explainable Time Series Classification

Understanding how a predictive model makes decisions is an increasingly important area of machine learning research. This is particularly critical for Time Series Classification (TSC) tasks, which have become pervasive across various domains like healthcare and sports science. In these applications, users need to understand the specific parts of the time series data that influence the model's classification outcome. For instance, imagine a physiotherapy exercise where a patient receives feedback on their performance (classification). Ideally, the feedback should not just say "correct" or "incorrect", but also explain which parts of the movement contain errors (explanation). This information allows the patient to take targeted corrective actions and improve their technique. The rise of explainable machine learning has led to numerous techniques bridging the gap between algorithms and users. These techniques generate explanations for many machine learning models, including TSC models. However, the inconsistency in their explanation outputs can be confusing to users and reduce user trust, especially in human-related domains like healthcare and sports analytics. While there is an abundance of new explanation methods, choosing the best one for a specific problem and dataset remains a challenge. This is particularly true for time series data, where the "ground truth" might be difficult for non-experts to grasp. Automated, computer-aided methods for evaluating and comparing explanations can be invaluable in filtering out problematic techniques and identifying those most trustworthy. In this thesis, we study and develop computational approaches to robustly assess attribution-based explanation methods for the TSC task. Attribution-based (or saliency-based) explanations account for the majority of the available explanation methods for this task. Saliency-based explanation considers each time point as a feature and assigns an importance weight for each time point. This type of explanation is often visualized by a heatmap overlaying the original data, making it visually friendly and helpful in pointing out the critical time points for the TSC algorithms. This thesis contributes one of the first explorations on computational evaluation of explanation methods. This exploration investigates the use of computational approaches to evaluate the informativeness of explanation methods. We extract importance weights for each point in the time series based on popular explanation methods and use these weights to perturb specific parts of the time series and measure the impact on classification accuracy. By this data perturbation process, we show that explanations that highlight discriminative sections of the time series lead to significant changes in classification accuracy, enabling the objective quantification and ranking of different explanations. We contribute a robust Model-Agnostic Explanation Evaluation framework (AMEE) that computationally assesses and ranks explanation methods for the TSC task. AMEE is an overhaul of our exploratory work, upgrading the capacity of the framework to be more robust by adding more perturbation methods that have defined specifications. We also introduce a standardized evaluation measure (Explanation Power) that is comparable across different explanation methods, referee classifiers, and datasets. Our experiments show a high agreement of the Explanation Power (measured by AMEE) in the synthetic datasets with the Oracle explanation (ground truth for each time point) and in selected real datasets with the expert explanations (ground truth provided by domain experts). We showcase two case studies using IMU sensors and video time series data. Our case studies show AMEE's ability to pinpoint the important data segments relevant to the specific classes, with domain experts validating the accuracy of these results.