Theory & literature

Calibrated Explanations builds on a multi-year research program spanning binary and multiclass classification, probabilistic and interval regression, and uncertainty-aware interpretation tooling. Use this page to locate published papers, benchmark references, and funding acknowledgements.

Publication highlights

Peer-reviewed work

  • Trustworthy explanations dissertation (2023): Introduces the decision-support principles behind calibrated explanations and documents the evaluation scaffolding used for subsequent studies. Accessible via the Jönköping University repository.

  • Expert Systems with Applications (2024): Binary-classification study with calibrated factual and alternative rules across 25 benchmarks. DOI: 10.1016/j.eswa.2024.123154.

  • Machine Learning journal (2025): Regression and probabilistic regression extension with interval guarantees mirroring the classification workflow. DOI: 10.1007/s10994-024-06642-8.

  • xAI 2024 proceedings: Conditional calibrated explanations for fairness-aware deployments. DOI: 10.1007/978-3-031-63787-2_17.

  • PMLR 230 (2024): Multiclass calibration analysis with empirical guarantees and tutorial material. Proceedings and artefacts available through the PMLR volume and paper/slides (MLResearch assets).

  • Annals of Mathematics and Artificial Intelligence (2023): First study that links calibration quality to explanation reliability. DOI: 10.1007/s10472-023-09837-2.

Preprints and in-flight studies

  • Ensured explanations (2024): Investigates strategies that actively reduce epistemic uncertainty in the generated alternatives. Preprint: arXiv:2410.05479.

  • Fast calibrated explanations (2024): Plugin-oriented acceleration of rule generation while preserving uncertainty guarantees. DOI: 10.1007/978-3-032-08333-3_16.

Benchmarks and evaluations

Calibrated Explanations has been stress-tested on 25 publicly available benchmark datasets for binary classification, with the same evaluation harness reused for regression and probabilistic extensions. The replication packages linked from the journal papers include dataset manifests, metric definitions, plots, and supporting notebooks so you can compare against your own workloads.

Funding and acknowledgements

The research has been funded by the Swedish Knowledge Foundation through the Knowledge Intensive Product Realization SPARK environment at Jönköping University. Grants include AFAIR (20200223), ETIAI (20230040), PREMACOP (20220187), and the Industrial Graduate School in Digital Retailing (INSiDR, 20160035). The project also thanks contributors for engineering support, including release automation, regression tooling, and third-party libraries such as crepes, venn-abers, ConformaSight, and scikit-learn components that ship with the distribution.

Keep exploring