Logo: to the web site of Uppsala University

Black-box classifiers are being deployed in high-stakes domains even though their lack oftransparency makes it difficult to trust and understand their predictions. The concept Rule-Based Model (RBM) mirror was implemented and evaluated as a framework for interpreting thedecision logic of black-box classifiers. The mirror was constructed using R.ROSETTA andtrained on relabeled data made of predictions from Multi-layer Perceptron (MLP) and SequenceU-Net classifiers. The RBM mirror was evaluated on the three classifiers trained on datasetscovering glioma diagnostics, e-mail spam classification and HPAIV protein sequences. The mirror achieved training accuracies between 92% and 96% and test accuracies between 91%and 95% on the neural network's predictions, confirming that it successfully captured thedecision patterns of the black-box models, including their misclassifications.

The RBM mirror was evaluated alongside established XAI methods, such as LIME and SHAP,to demonstrate the capabilities of a comprehensive approach to transparency. Throughvisualizing rule networks and a misclassification heatmap, the mirror revealed combinatorialfeature synergies and potential systematic error patterns that feature importance score alonecannot capture. By translating the internalized logic of black-box models into discrete IF-THENrules, the RBM mirror provides a transparent and verifiable decision structure that movesbeyond traditional post-hoc explanations.