Disclaimer
XAI4TSC is a research framework. Please read the points below before relying on its outputs for scientific conclusions.
Implementation correctness
Many of the models, explainers, and metrics shipped with XAI4TSC are clean-room re-implementations of methods from the research literature. These implementations have not been verified by the original authors and may differ from the source papers or any reference code — in subtleties of normalisation, default parameters, or numerical detail that can change results.
Where XAI4TSC instead wraps a third-party library (for example, Captum for feature attributions or Quantus for evaluation metrics), the behaviour and caveats of that library apply, and those wrapped implementations carry their own assumptions that are likewise not guaranteed to match a method’s original description.
Evaluation metrics are interpretations
Explanation-quality metrics are an empirical, and sometimes contested, proxy for what a “good” explanation is. A metric measures what its implementation computes, which may not coincide with what the original authors intended it to capture, and different metrics can disagree. High coverage of a metric’s code, or a confident-looking number, does not by itself establish that an explanation is faithful or useful.
Using XAI4TSC responsibly
When using XAI4TSC for research or in publications, we recommend that you:
Cite the original papers for every model, explainer, and metric you use, not just XAI4TSC.
State which implementation you used (XAI4TSC’s re-implementation versus a wrapped library) and its version.
Validate against a reference implementation where one is available, and report any discrepancies.
Interpret scores in context — compare methods under identical settings rather than reading absolute values as ground truth.
If you find a discrepancy between an XAI4TSC implementation and the method it is based on, please open an issue or a pull request (see CONTRIBUTING.md). Corrections and clarifications are very welcome.