Open Access

Table 5

Held-out test set performance with BCa bootstrap 95% CIs.

Model Accuracy (%) Macro F1 (%) MCC
XGBoost 99.89 [99.63, 99.96] 99.89 [99.70, 99.96] 0.999
RF 99.85 [99.59, 99.93] 99.85 [99.63, 99.96] 0.998
SoftCBM 99.04 [98.56, 99.33] 99.04 [98.61, 99.37] 0.988
HardCBM-Cal 96.48 [95.67, 97.07] 96.46 [95.73, 97.09] 0.957
CEM 96.70 [95.93, 97.26] 96.69 [95.98, 97.30] 0.960
MLP 95.04 [94.15, 95.78] 95.00 [94.18, 95.76] 0.940
HardCBM 94.19 [93.22, 94.96] 94.12 [93.22, 94.94] 0.930
HardCBM-Lin 88.85 [87.59, 89.96] 88.65 [87.46, 89.83] 0.866

Notes. BCa bootstrap with 10 000 resamples. Brackets show 95% confidence intervals. The held-out predictions summarized in this table are drawn from models trained with the identical hyperparameters reported in Sect. 3.3; the small differences between cross-validation means in Table 4 and the bootstrap means here (e.g., HardCBM 94.41% vs. 94.19%) reflect the CV-versus-held-out test split, not model differences.

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.