Table 2.

Results of CNN-DDI using different features

Feature	ACC	AUPR	AUC	F1	Precision	Recall
T	0.7915	0.8470	0.9953	0.6099	0.6932	0.5716
P	0.7820	0.8381	0.9952	0.5805	0.6822	0.5364
E	0.6580	0.7098	0.9897	0.3344	0.4419	0.2957
C	0.8702	0.9139	0.9966	0.7421	0.7994	0.7125
T + P	0.8227	0.8898	0.9969	0.6778	0.7589	0.6375
T + E	0.8242	0.8712	0.9956	0.6360	0.7373	0.5849
T + C	0.8792	0.9185	0.9960	0.7627	0.8167	0.7405
P + E	0.8255	0.8747	0.9958	0.6227	0.7130	0.5781
P + C	0.8796	0.9179	0.9961	0.7440	0.7955	0.7485
E + C	0.8496	0.8895	0.9948	0.6928	0.7726	0.6488
T + P + E	0.8243	0.8690	0.9947	0.6489	0.7332	0.6063
T + P + C	0.8797	0.9199	0.9960	0.7490	0.8164	0.7232
T + E + C	0.8539	0.8899	0.9933	0.6938	0.7726	0.6539
P + E + C	0.8559	0.8919	09,939	0.6845	0.7575	0.6485
T + P + E + C	0.8871	0.9251	0.9980	0.7496	0.8556	0.7220

The bold values indicate the result of CNN_DDI with four types of features. So it can be concluded that the drug category is effective as a new feature type and multiple features can imporve the performanced of CNN-DDI