Deep Learning-Based Prediction of Physical Stability considering Class Imbalance for Amorphous Solid Dispersions

Table 3

The performance comparison of models applying various sampling techniques to the lightGBM [9].

		Without sampling	Oversampling		Undersampling				Hybrid sampling
			SMOTE	ADASYN	TL	ENN	OSS	CNN	Smote+ENN	Smote+TL
Accuracy	Training set	0.99	1.0	1.0	1.0	1.0	1.0	0.364	1.0	1.0
	5-fold cross-validation
	Test set	0.9	0.9	0.85	0.75	0.6	0.8	0.05	0.85	0.85
Precision	Training set	0.991	1.0	1.0	1.0	1.0	1.0	0.132	1.0	1.0
	5-fold cross-validation
	Test set	0.86	0.86	0.83	0.709	0.734	0.807	0.002	0.866	0.83
Recall	Training set	0.99	1.0	1.0	1.0	1.0	1.0	0.364	1.0	1.0
	5-fold cross-validation
	Test set	0.9	0.9	0.85	0.75	0.6	0.8	0.05	0.85	0.85
F1-score	Training set	0.99	1.0	1.0	1.0	1.0	1.0	0.194	1.0	1.0
	5-fold cross-validation
	Test set	0.878	0.878	0.832	0.729	0.653	0.788	0.005	0.856	0.832