Synthetic Multi-Label Time Series Classification Dataset for XAI Evaluation

This dataset provides a synthetic, multi-label time series classification benchmark designed to support the evaluation of eXplainable Artificial Intelligence (XAI) methods using ground-truth explanations. Unlike existing synthetic datasets, which typically cover only single-label settings, this dataset addresses multi-label time series classification. Discriminative patterns are located in both the time and frequency domains, enabling evaluation of explanation methods across time, frequency, and time-frequency representations. The dataset contains 4,000 univariate time series samples of 500 time steps each. Every sample carries zero or more of three possible class labels, defined by the presence of class-specific frequency windows. The three labels may co-occur, resulting in eight possible label combinations, each represented equally in the dataset. Co-occurring class-specific frequency features overlap by up to 50% of their frequency windows, allowing evaluation of explanation methods under feature overlap. Non-discriminative features and Gaussian noise are added as distractors to simulate irrelevant signal content. Alongside the dataset, we provide trained models of common time series classification architectures, supporting reproducible benchmarking and evaluation of XAI methods. Code used to generate the data and train the models is available on GitHub: https://github.com/TimeXAIgroup/multilabel-ts-xai-benchmark Acknowledgements This work was supported by the German Federal Ministry of Research, Technology and Space (BMFTR) through ExperTeam4KI (grant no. 16IS24063). We gratefully acknowledge support from the hessian.AI Service Center (funded by the BMFTR, grant no. 16IS22091) and the hessian.AI Innovation Lab (funded by the Hessian Ministry for Digital Strategy and Innovation, grant no. S-DIW04/0013/003).