1 September 2026 in SSRN Electronic Journal
Evaluating explanations is an essential task in the field of eXplainable Artificial Intelligence (XAI) for time series. One approach to evaluate explanations is using ground-truth explanations that describe which features are expected to be relevant. However, defining such ground truth is particularly challenging for real-world time series classification tasks, where the temporal and spectral regions responsible for a prediction are rarely known. While some synthetic datasets provide ground-truth explanations, they cover only a limited range of time series properties, such as simple temporal patterns, single-label classification, and non-overlapping explanations. Existing datasets do not address multi-label time series classification, leaving this setting without a dedicated benchmark.To address this gap, we present a novel synthetic multi-label time series classification dataset designed to support the evaluation of XAI methods using ground-truth explanations. This dataset provides a controlled benchmark setting, in which multiple labels and their corresponding discriminative patterns can co-occur within a single time series sample. Discriminative patterns are located in both the time and frequency domains, enabling the evaluation of XAI methods across time, frequency, and time-frequency domains.The dataset contains 4000 univariate time series samples of 500 time steps each. Each sample contains between zero and three possible class labels defined by the presence of class-specific frequency windows. The three class labels may co-occur within a sample, resulting in eight possible combinations, each represented equally across 500 samples. In addition, the samples contain Gaussian noise and non-discriminative frequency components as distractors to simulate irrelevant signal components. Co-occurring class-specific frequency features overlap by 25% to 50% of the frequency windows, allowing the evaluation of explanation methods under feature overlap.In addition to the dataset, we provide trained models covering common time series classification architectures in order to support benchmarking and reproducible evaluation of XAI methods. The dataset and trained models are available on Zenodo, and the code used to generate the data and to train the models can be accessed via GitHub.