Table 4
Model performance (five-fold CV on NCV = 15 300).
| Model | Acc. (%) | Macro F1 (%) | MCC | Interp. |
|---|---|---|---|---|
| XGBoost | 99.81 ± 0.11 | 99.81 ± 0.11 | 0.998 | – |
| Random forest | 99.79 ± 0.09 | 99.79 ± 0.09 | 0.998 | – |
| SoftCBM | 99.12 ± 0.29 | 99.12 ± 0.29 | 0.989 | Partial |
| CEM | 97.13 ± 0.36 | 97.13 ± 0.37 | 0.965 | Partial |
| HardCBM-Cal | 96.97 ± 0.47 | 96.96 ± 0.47 | 0.964 | Partiald |
| MLP | 95.84 ± 0.29 | 95.83 ± 0.29 | 0.950 | – |
| HardCBM | 94.41 ± 0.36 | 94.37 ± 0.38 | 0.933 | Full |
| HardCBM-Lin | 90.67 ± 0.85 | 90.59 ± 0.87 | 0.888 | Full |
Notes. “Interp.” indicates interpretability level: “Full” = predictions traceable through named scalar concepts with intervention capability; “Partial” = concepts identifiable but not directly interventionable; “-” = black-box. MCC = Matthews Correlation Coefficient (range [−1,1]; >0.9 indicates excellent agreement). AUC-ROC values (one-vs-rest, macro) are >0.99 for all models and omitted for brevity; per-class AUC-ROC is provided in the Zenodo archive (see Data availability). (d)HardCBM-Cal uses cross-input calibrators (each receiving all 12 concepts) that permit inter-concept information flow; its interpretability is weaker than HardCBM’s strict per-concept independence.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.