Leveraging augmentation techniques for tasks with unbalancedness within the financial domain: a two-level ensemble approach

EPJ Data Science

Table 13 Classification results for the bankruptcy dataset in USA

Evaluation metric	Naive Bayes	SGD	KNN	MLP	Random forest	SVM
Without augmentation
F1 score macro	0.3416	0.5779	0.3917	0.3855	0.4059	0.3478
F1 score micro	0.5202	0.6278	0.5398	0.5381	0.5482	0.521
Precision minority class	0	0.7791	0.8651	0	0.95	0
Recall minority class	0	0.3195	0.0493	0.0443	0.0632	0.0067
Baseline: min-max approach
F1 score macro	0.347	0.4792	0.4797	0.4819	0.359	0.343
F1 score micro	0.5202	0.5772	0.5816	0.5312	0.5277	0.5202
Precision minority class	0	0.7465	0.8588	0.5157	0	0
Recall minority class	0.0061	0.1637	0.1582	0.2434	0.0157	0.0017
Baseline: mean-std approach
F1 score macro	0.3416	0.5315	0.3429	0.3781	0.372	0.35
F1 score micro	0.5202	0.5953	0.5052	0.5344	0.5331	0.5227
Precision minority class	0	0.7194	0	0	0.9762	0
Recall minority class	0	0.2693	0.01	0.0367	0.0288	0.0083
SMOTE approach
F1 score macro	0.7136	0.7051	0.9246	0.7693	0.9823	0.5228
F1 score micro	0.7157	0.7093	0.9253	0.7715	0.9825	0.5272
Precision minority class	0.7113	0.7231	0.8639	0.77	0.9804	0.5051
Recall minority class	0.6863	0.6507	1	0.7436	0.9823	0.5016
SMOTE-clustering approach
F1 score macro	0.7111	0.7442	0.9179	0.7428	0.9875	0.5261
F1 score micro	0.7132	0.7458	0.9186	0.7452	0.9875	0.5305
Precision minority class	0.7097	0.7244	0.8537	0.7534	0.9852	0.5094
Recall minority class	0.6813	0.7715	1	0.699	0.99	0.495