Investigating ADR mechanisms with Explainable AI: a feasibility study with knowledge graph mining

Table 7 Robustness evaluation of our classifiers

Data set	Accuracy	AUC	F1-score \(\oplus\)
\(DILI^{shuffled}\)	0.52	0.49	0.36
\(DILI^{random \ominus _1}\)	0.92	0.91	0.89
\(DILI^{random \ominus _2}\)	0.92	0.91	0.89
\(DILI^{random \ominus _3}\)	0.93	0.92	0.91
\(DILI^{random \ominus _4}\)	0.93	0.92	0.90
\(DILI^{random \ominus _5}\)	0.92	0.91	0.90
\(SCAR^{shuffled}\)	0.63	0.51	0.26
\(SCAR^{random \ominus _1}\)	0.93	0.89	0.86
\(SCAR^{random \ominus _2}\)	0.94	0.90	0.88
\(SCAR^{random \ominus _3}\)	0.93	0.90	0.86
\(SCAR^{random \ominus _4}\)	0.92	0.89	0.85
\(SCAR^{random \ominus _5}\)	0.93	0.89	0.86

\(\cdot ^{shuffled}\) corresponds to an experiment where class labels (i.e., \(\oplus\) or \(\ominus\)) are randomly affected to drugs. \(\cdot ^{random \ominus _{i}}\) correspond to experiments where negative examples (i.e., \(\ominus\)) are replaced by drugs randomly picked in the knowledge graph. Indices i from 1 to 5 refer to 5 different draws

ISSN: 1472-6947