IMOVNO+: A Regional Partitioning and Meta-Heuristic Ensemble Framework for Imbalanced Multi-Class Learning
Authors: Soufiane Bacha, Laouni Djafri, Sahraoui Dhelim, Huansheng Ning
Organizations: School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China. · Key Laboratory of Xinjiang Coal Resources Green Mining, Ministry of Education, Urumqi,830023,China · Department of Mathematics, Ibn Khaldoun University,Tiaret, 14000,Algeria · LIM Laboratory of Informatics and Mathematics, Ibn Khaldoun University,Tiaret,Algeria · School of Computing, Dublin City University, Ireland
Class imbalance, overlap, and noise degrade data quality, reduce model reliability, and limit generalization. Although widely studied in binary classification, these issues remain underexplored in multi-class settings, where complex inter-class relationships make minority-majority structures unclear and traditional clustering fails to capture distribution shape. Approaches that rely only on geometric distances risk removing informative samples and generating low-quality synthetic data, while binarization approaches treat imbalance locally and ignore global inter-class dependencies. At the algorithmic level, ensembles struggle to integrate weak classifiers, leading to limited robustness. This paper proposes IMOVNO+ (IMbalance-OVerlap-NOise+ Algorithm-Level Optimization), a two-level framework designed to jointly enhance data quality and algorithmic robustness for binary and multi-class tasks. At the data level, first, conditional probability is used to quantify the informativeness of each sample. Second, the dataset is partitioned into core, overlapping, and noisy regions. Third, an overlapping-cleaning algorithm is introduced that combines Z-score metrics with a big-jump gap distance. Fourth, a smart oversampling algorithm based on multi-regularization controls synthetic sample proximity, preventing new overlaps. At the algorithmic level, a meta-heuristic prunes ensemble classifiers to reduce weak-learner influence. IMOVNO+ was evaluated on 35 datasets (13 multi-class, 22 binary). Results show consistent superiority over state-of-the-art methods, approaching 100% in several cases. For multi-class data, IMOVNO+ achieves gains of 37-57% in G-mean, 25-44% in F1-score, 25-39% in precision, and 26-43% in recall. In binary tasks, it attains near-perfect performance with improvements of 14-39%. The framework handles data scarcity and imbalance from collection and privacy limits.
Figures & tables
Fig. 1: Flowchart of the proposed framework
LiOptimal=LiInit(j∗)
Algorithm 6 Optimized Ensemble Learning (OEL)
Datasets
Features
Instances
IR
No. of classes
Contraceptive
9
1473
1.89
3
Dermatology
34
366
5.55
6
Balance
4
625
5.88
3
New-thyroid
5
215
5
3
Zoo
16
101
10.25
7
Yeast
8
1484
92.6
10
TABLE 1: Description of multi-class imbalanced datasets.
Datasets
Instances
Features
IR
Yeast1
1484
8
2.46
Ecoli1
336
8
3.36
Yeast-1_vs_7
459
7
14.30
Yeast2vs8
482
9
23.10
Glass2
214
9
11.59
Abalone9vs18
731
9
16.40
TABLE 2: Description of binary-class imbalanced datasets.
Methods / References
Parameter Values
SMOTE-CDNN [9]
k_neighbors=21
ECDNN [ 63 ] [ 64 ]
k_neighbors=21,n (vote hyper-parameter)=2
SAMME.C2 [ 18 ]
Number of iterations = 200
CPS-3WS [ 65 ]
α=0.4,β=0.5,k_neighbors=5, Base classifier: decision tree
Counterfactual SMOTE [ 66 ]
Strategy: flexible, Number of optimization steps = 3 Number of iterations = 3 Base classifier: random forest
MLOS [ 67 ]
α=0.4,k_neighbors=5, Base classifier: SVM
TABLE 3: Overview of the parameter settings.
Fig. 2: Conditional region partition with lower imbalanced ratio ( IR ≤ 3 ).
Fig. 3: Conditional region partition with medium imbalanced ratio (3 ≤ IR ≤ 9).
Fig. 4: Conditional region partition with higher imbalance ratio (IR > 9).
IR Group
Datasets
OR (Before)
OR (After)
Reduction
IR <3
Contraceptive
29.12
16.27
44.13
vehicle
36.73
25.14
31.54
vertebral
10.78
9.16
15.00
3<IR<9
balance
0.46
0.39
16.28
bioconcentration
27.56
11.13
59.61
glass
26.23
24.59
6.25
TABLE 4: Quantitative evaluation of overlap reduction based on OR.
Fig. 5: Overlapping removal on the Contraceptive dataset (IR ≤ 3).
Fig. 6: Overlapping removal on the Bioconcentration dataset (3 ≤ IR ≤ 9).
Fig. 7: Overlapping removal on the Winequality-red dataset (IR > 9).
Fig. 8: Balancing data on Contraceptive dataset(IR ≤ 3).
Fig. 9: Balancing data on Bioconcentration dataset(3 ≤ IR ≤ 9).
Fig. 10: Balancing data on winequality-red dataset (IR > 9)
Dataset (multi-class)
Framework components
Multi-class Metrics
Dataset (binary)
Binary-class Metrics
Acc
Prec
Recall
G-mean
F-score
Acc
Prec
Recall
G-mean
F-score
IR < 3
Contraceptive
CCRP+SOR+OEL
80.23 ± 0.0092
81.70 ± 0.0091
79.80 ± 0.0104
79.56 ± 0.0103
80.34 ± 0.0094
yeast1
94.58 ± 0.0042
97.32 ± 0.0041
94.23 ± 0.0050
95.72 ± 0.0034
94.69 ± 0.0046
CCRP+SOR+(OVO+basic-smote)+OEL
54.56 ± 0.0087
54.38 ± 0.0094
53.08 ± 0.0086
50.66 ± 0.0092
52.29 ± 0.0088
65.81 ± 0.0104
50.52 ± 0.0155
65.54 ± 0.0421
64.60 ± 0.0160
55.94 ± 0.0202
CCRP+SOR+CE+EL
84.85 ± 0.0499
84.23 ± 0.0857
84.85 ± 0.0499
73.93 ± 0.1600
82.33 ± 0.0697
90.77 ± 0.0123
86.39 ± 0.0200
97.71 ± 0.0093
90.35 ± 0.0137
91.52 ± 0.0101
CCRP+SOR+CE+OEL (IMOVNO+)
95.26 ± 0.0083
95.58 ± 0.0066
95.26 ± 0.0083
95.10 ± 0.0091
95.23 ± 0.0084
95.06 ± 0.0027
98.12 ± 0.0031
91.91 ± 0.0064
94.99 ± 0.0028
94.88 ± 0.0029
TABLE 5: Component-wise ablation study of the proposed framework.
Dataset
DT-E-EVRS (2023) ) [ 17 ]
SVM-E-EVRS (2023) [ 17 ]
CPSQSVM (2025) [ 73 ]
IMOVNO+ (Proposed)
F1 -score( % )
G-mean( % )
F1 -score( % )
G-mean( % )
Accuracy %
F1 -score( % )
Accuracy %
F1 -score( % )
G-mean( % )
Contraceptive
53.90 ± 0.041
53.80 ± 0.042
53.90 ± 0.033
54.50 ± 0.033
54.67 ± 0.22
53.71 ± 0.20
95.26 ± 0.0083
95.23 ± 0.0084
95.10 ± 0.0091
Balance
84.90 ± 0.085
89.70 ± 0.034
81.20 ± 0.061
88.50 ± 0.066
90.59 ± 0.67
71.44 ± 0.91
99.64 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
New-thyroid
94.90 ± 0.053
95.20 ± 0.060
95.60 ± 0.030
96.30 ± 0.036
96.09 ± 0.37
93.81 ± 0.76
99.50 ± 0.0026
99.50 ± 0.0027
99.49 ± 0.0027
Zoo
92.80 ± 0.110
70.00 ± 0.458
91.80 ± 0.136
78.90 ± 0.445
96.69 ± 0.91
89.11 ± 2.71
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
Yeast
56.30 ± 0.151
70.40 ± 0.039
59.80 ± 0.061
59.20 ± 0.223
58.15 ± 0.33
52.24 ± 0.97
91.77 ± 0.0123
91.53 ± 0.0128
90.62 ± 0.0154
TABLE 6: Comparison analysis with published state-of-the-art results on multi-class imbalanced datasets using 5-fold cross-Validation.
Dataset
MDGP-Forest (2025) [ 74 ]
GDHS (2025) [ 75 ]
FEHC (2024) [ 76 ]
IMOVNO+ (Proposed)
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Vehicle
81.03 ± 0.0261
87.32 ± 0.0181
68.93 ± 0.0219
69.12 ± 0.0246
68.85 ± 0.0227
65.76 ± 0.0305
73.24 ± 0.0148
74.54 ± 0.0123
73.50 ± 0.0139
82.84 ± 0.0100
79.41 ± 0.0060
77.00 ± 0.0068
70.59 ± 0.0105
Yeast
58.35 ± 0.0614
73.69 ± 0.0412
45.14 ± 0.0314
51.60 ± 0.0228
45.64 ± 0.0260
38.28 ± 0.1126
47.07 ± 0.0154
55.86 ± 0.0145
48.08 ± 0.0147
93.20 ± 0.0093
91.81 ± 0.0121
91.53 ± 0.0128
90.62 ± 0.0154
Balance
93.92 ± 0.0282
96.05 ± 0.0203
61.58 ± 0.0302
59.67 ± 0.0531
59.47 ± 0.0386
50.64 ± 0.0960
64.40 ± 0.0097
61.55 ± 0.0199
60.78 ± 0.0136
99.65 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
New-thyroid
96.54 ± 0.0265
97.12 ± 0.0299
89.98 ± 0.0333
93.62 ± 0.0317
91.39 ± 0.0290
93.33 ± 0.0346
90.51 ± 0.0163
95.24 ± 0.0172
92.30 ± 0.0159
99.52 ± 0.0025
99.50 ± 0.0026
99.50 ± 0.0027
99.49 ± 0.0027
Glass
76.36 ± 0.0552
84.82 ± 0.0397
67.62 ± 0.0726
74.06 ± 0.0435
66.72 ± 0.0580
69.43 ± 0.0534
65.73 ± 0.0250
74.22 ± 0.0336
65.22 ± 0.0289
93.35 ± 0.0145
91.85 ± 0.0155
91.68 ± 0.0164
90.06 ± 0.0395
TABLE 7: Comparison analysis with published state-of-the-art results on multi-class imbalanced datasets using 5-Fold Cross-Validation.
Dataset
SAMME.C2 (2024) [ 18 ]
ECDNN (2023) [ 63 ]
SMOTE-CDNN (2023) [ 64 ]
IMOVNO+ (Proposed)
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Contraceptive
53.81 ± 0.0077
52.69 ± 0.0085
52.85 ± 0.0083
51.09 ± 0.0088
51.13 ± 0.0095
50.32 ± 0.0084
50.48 ± 0.0088
49.04 ± 0.0093
52.58 ± 0.0063
52.65 ± 0.0071
50.35 ± 0.0071
51.80 ± 0.0073
95.58 ± 0.0066
95.26 ± 0.0083
95.23 ± 0.0084
95.10 ± 0.0091
Balance
79.97 ± 0.0179
83.42 ± 0.0220
80.83 ± 0.0189
81.06 ± 0.0359
66.99 ± 0.0278
65.90 ± 0.0089
65.17 ± 0.0153
31.31 ± 0.0943
79.10 ± 0.0073
88.71 ± 0.0069
80.10 ± 0.0113
88.53 ± 0.0073
99.65 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
Dermatology
37.60 ± 0.0039
47.65 ± 0.0027
39.15 ± 0.0039
03.16 ± 0.0025
87.06 ± 0.0099
87.73 ± 0.0084
86.68 ± 0.0085
86.70 ± 0.0106
96.46 ± 0.0066
96.35 ± 0.0060
96.22 ± 0.0064
96.07 ± 0.0066
98.62 ± 0.0012
98.55 ± 0.0011
98.54 ± 0.0012
98.50 ± 0.0012
Zoo
65.38 ± 0.0366
74.11 ± 0.0199
67.54 ± 0.0279
24.03 ± 0.0636
78.76 ± 0.0213
81.10 ± 0.0224
78.79 ± 0.0198
37.67 ± 0.0637
83.07 ± 0.0166
86.09 ± 0.0154
83.38 ± 0.0158
55.22 ± 0.0406
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
Yeast
35.91 ± 0.0201
35.35 ± 0.0161
33.44 ± 0.0168
07.35 ± 0.0134
60.20 ± 0.0097
56.36 ± 0.0071
57.15 ± 0.0079
29.72 ± 0.0089
45.54 ± 0.0129
56.34 ± 0.0036
44.19 ± 0.0056
44.24 ± 0.0331
93.20 ± 0.0093
91.81 ± 0.0121
91.53 ± 0.0128
90.62 ± 0.0154
TABLE 8: Comparison analysis with published state-of-the-art and implemented results on multi-Class imbalanced datasets using 5-fold cross-validation.
Dataset
OVO-MLOS (2025) [ 67 ]
OVO-Counterfactual SMOTE (2025) [ 66 ]
OVO-CPS-3WS (2024) [ 65 ]
IMOVNO+ (Proposed)
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Contraceptive
53.38 ± 0.0078
51.99 ± 0.0066
52.22 ± 0.0071
50.09 ± 0.0073
50.33 ± 0.0095
49.38 ± 0.0076
49.53 ± 0.0082
47.54 ± 0.0093
45.29 ± 0.0065
44.92 ± 0.0051
44.76 ± 0.0057
42.61 ± 0.0085
95.58 ± 0.0066
95.26 ± 0.0083
95.23 ± 0.0084
95.10 ± 0.0091
Balance
60.40 ± 0.0016
65.46 ± 0.0017
62.81 ± 0.0017
09.88 ± 0.0002
59.10 ± 0.0033
61.87 ± 0.0055
60.40 ± 0.0042
09.51 ± 0.0006
58.32 ± 0.0071
57.46 ± 0.0089
57.79 ± 0.0079
05.21 ± 0.0576
99.65 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
Dermatology
97.32 ± 0.0019
97.30 ± 0.0021
97.22 ± 0.0021
97.11 ± 0.0025
97.92 ± 0.0036
97.61 ± 0.0043
97.64 ± 0.0043
97.42 ± 0.0050
94.92 ± 0.0077
94.37 ± 0.0158
94.38 ± 0.0120
94.01 ± 0.0179
98.62 ± 0.0012
98.55 ± 0.0011
98.54 ± 0.0012
98.50 ± 0.0012
Zoo
84.82 ± 0.0261
86.81 ± 0.0169
84.97 ± 0.0216
57.73 ± 0.0786
50.33 ± 0.0095
49.38 ± 0.0076
49.53 ± 0.0082
47.54 ± 0.0093
80.46 ± 0.0361
82.27 ± 0.0445
80.91 ± 0.0370
49.83 ± 0.1329
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
Yeast
59.75 ± 0.0122
54.67 ± 0.0060
55.88 ± 0.0064
27.61 ± 0.0113
55.79 ± 0.0155
51.38 ± 0.0078
52.07 ± 0.0085
69.67 ± 0.0055
61.79 ± 0.0601
65.57 ± 0.0512
62.35 ± 0.0542
52.25 ± 0.0984
93.20 ± 0.0093
91.81 ± 0.0121
91.53 ± 0.0128
90.62 ± 0.0154
TABLE 9: Comparison of implemented results on multi-class datasets using binarization and 5-fold cross-validation.
Dataset
OVO-HSCF (2025) [ 68 ]
OVO-GDDSAD (2025) [ 69 ]
OVO-ISMOTE (2025) [ 77 ]
IMOVNO+ (Proposed)
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
Contraceptive
45.29 ± 0.0065
44.92 ± 0.0051
44.76 ± 0.0057
42.61 ± 0.0085
47.01 ± 0.0519
43.38 ± 0.0108
34.65 ± 0.0258
24.18 ± 0.0708
52.81 ± 0.0118
52.48 ± 0.0129
52.46 ± 0.0123
51.549 ± 0.0145
95.58 ± 0.0066
95.26 ± 0.0083
95.23 ± 0.0084
95.10 ± 0.0091
Balance
58.32 ± 0.0071
57.46 ± 0.0089
57.79 ± 0.0079
13.14 ± 0.0455
78.15 ± 0.0104
86.86 ± 0.0164
78.93 ± 0.0127
86.64 ± 0.0166
59.84 ± 0.0042
60.43 ± 0.0042
60.07 ± 0.0037
15.40 ± 0.0560
99.65 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
99.64 ± 0.0007
Dermatology
95.25 ± 0.0098
94.19 ± 0.0138
94.32 ± 0.0139
93.56 ± 0.0184
94.88 ± 0.0079
93.71 ± 0.0110
93.87 ± 0.0099
93.17 ± 0.0120
97.82 ± 0.0061
97.59 ± 0.0060
97.57 ± 0.0064
97.40 ± 0.0067
98.62 ± 0.0012
98.55 ± 0.0011
98.54 ± 0.0012
98.50 ± 0.0012
Zoo
83.13 ± 0.0347
85.55 ± 0.0287
83.29 ± 0.0328
52.18 ± 0.0852
78.08 ± 0.0382
78.71 ± 0.0424
75.23 ± 0.0410
50.97 ± 0.1025
86.60 ± 0.0162
90.62 ± 0.0111
89.33 ± 0.0130
67.55 ± 0.0493
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
100.00 ± 0.0000
Yeast
50.88 ± 0.0250
49.84 ± 0.0205
49.16 ± 0.0208
24.66 ± 0.0350
45.55 ± 0.0187
48.96 ± 0.0181
45.48 ± 0.0160
27.84 ± 0.0386
57.71 ± 0.0120
59.35 ± 0.0118
57.78 ± 0.0107
36.88 ± 0.0306
93.20 ± 0.0093
91.81 ± 0.0121
91.53 ± 0.0128
90.62 ± 0.0154
TABLE 10: Comparison of implemented results on multi-class datasets using binarization and 5-fold cross-validation.
Dataset
SIMPOR (2023) [ 78 ]
Adaptive SV-Borderline SMOTE-SVM (2024) [ 79 ]
SVM-OMOS (2024) [ 80 ]
UBMD (2024) [ 81 ]
IMOVNO+ (Proposed)
Precision( % )
Recall( % )
F1 -score( % )
F1 -score( % )
G-mean( % )
F1 -score( % )
G-mean( % )
Recall( % )
F1 -score( % )
G-mean( % )
Accuracy( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
yeast1
72.70
70.40
71.50
79.00
68.30
58.9 ± 0.012
71.10 ± 0.010
78.84
63.32
75.00
95.06 ± 0.0027
98.12 ± 0.0031
91.91 ± 0.0064
94.88 ± 0.0029
94.99 ± 0.0028
Ecoli1
81.00
85.40
83.10
88.60
87.90
83.9 ± 0.022
89.9 ± 0.016
92.80
80.29
88.10
99.18 ± 0.0042
98.61 ± 0.0068
99.80 ± 0.0024
99.17 ± 0.0042
99.19 ± 0.0041
yeast-1_vs_7
77.90
65.50
71.00
50.20
69.70
30.8 ± 0.069
7.72 ± 0.077
78.67
26.61
73.92
96.98 ± 0.0052
95.05 ± 0.0099
99.26 ± 0.0078
97.07 ± 0.0049
96.94 ± 0.0054
Yeast2_vs_8
91.30
87.20
88.40
98.70
74.30
71.4 ± 0.099
93.90 ± 0.064
90.00
28.77
83.52
98.80 ± 0.0047
100.00 ± 0.0000
97.62 ± 0.0094
98.77 ± 0.0050
98.79 ± 0.0048
glass2
73.10
74.70
73.70
53.00
78.10
45.30 ± 0.106
83.50 ± 0.071
76.33
25.84
68.18
92.99 ± 0.0049
88.36 ± 0.0066
99.31 ± 0.0025
93.46 ± 0.0042
92.74 ± 0.0053
TABLE 11: Comparison analysis with published state-of-the-art and implemented results on binary imbalanced datasets using 5-fold cross-validation.
Dataset
DT-Boosting
LR-DC-Glow (2023) [ 82 ]
MLCC-IBC (2023) [ 83 ]
IMOVNO+ (Proposed)
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
F1 -score( % )
G-mean( % )
F1 -score( % )
G-mean( % )
Precision( % )
Recall( % )
F1 -score( % )
G-mean( % )
messidor_features
68.33 ± 0.0143
63.19 ± 0.0296
65.40 ± 0.0128
64.42 ± 0.0091
75.56
74.85
75.07
67.92
98.20 ± 0.0061
87.24 ± 0.0171
92.29 ± 0.0070
92.56 ± 0.0064
Pima
66.66 ± 0.0209
56.70 ± 0.0201
60.92 ± 0.0094
69.02 ± 0.0081
67.88
75.22
69.94
75.53
97.42 ± 0.0062
95.02 ± 0.0094
96.15 ± 0.0044
96.17 ± 0.0042
biodeg
74.84 ± 0.0146
73.51 ± 0.0250
73.95 ± 0.0118
79.96 ± 0.0107
78.72
83.87
84.07
88.61
97.89 ± 0.0015
92.70 ± 0.0067
95.20 ± 0.0037
95.30 ± 0.0035
vehicle3
60.66 ± 0.0349
29.91 ± 0.0175
39.44 ± 0.0204
52.33 ± 0.0164
71.32
81.97
68.80
82.05
95.00 ± 0.0078
94.74 ± 0.0068
94.84 ± 0.0057
94.82 ± 0.0057
Ecoli1
78.64 ± 0.0330
76.07 ± 0.0465
76.51 ± 0.0352
83.97 ± 0.0283
75.28
86.15
79.12
90.59
98.61 ± 0.0068
99.80 ± 0.0024
99.17 ± 0.0042
99.19 ± 0.0041
TABLE 12: Comparison analysis on binary imbalanced datasets using 5-fold cross-validation.
Fig. 11: Comparative analysis of accuracy and AUC scores across various multi-class imbalanced datasets.
Fig. 12: Experimental results of noise reduction for different noise thresholds (IR ≤ 3).
Fig. 13: Experimental results of noise reduction for different noise thresholds (3 ≤ IR ≤ 9).
Fig. 14: Experimental results of noise reduction for different noise thresholds (IR > 9).
Fig. 15: Comparative analysis using the strong classifiers.
Fig. 16: Critical difference diagram from the Nemenyi test comparing 10 rebalancing approaches across 13 datasets.
Class imbalance poses a significant challenge in classification, where existing methods such as SMOTE often generate low-quality synthetic samples in regions with noise or class overlap. We propose QC-SMOTE, a quality-controlled oversampling framework that estimates minority sample reliability using a composite neighbourhood trustworthiness score combining local density, safe-level, and isolation from the majority class. Synthetic candidates are generated using an IPQ-guided best-of-K strategy that evaluates midpoint purity and, when required, majority clearance, with allocation guided by sample reliability and boundary informativeness. Generation behaviour adapts across overlap--imbalance regimes, adjusting interpolation range and selection criteria to match local data geometry. Low-quality synthetic samples are replaced with original minority duplicates when neighbourhood purity falls below an adaptive threshold, providing graceful degradation by reverting to duplication in severely noisy regions. Experiments on 30 imbalanced datasets using repeated stratified cross-validation show that QC-SMOTE achieves the strongest average AUC-ROC and Macro F1 among the compared oversampling methods, with particularly clear gains under moderate and severe imbalance. These results demonstrate the importance of quality-aware, geometry-adaptive synthetic sampling for robust imbalanced classification.
Parth Upman, Shreyank N Gowda
School of Computer Science, University of Nottingham, Jubilee Campus, Wollaton Road, Nottingham, NG8 1BB, United Kingdom
Class imbalance complicates probabilistic classification because standard training objectives emphasize majority-class performance. Synthetic oversampling can reduce imbalance, but discrepancies between the synthetic and target minority distributions may bias the fitted classifier, especially because synthetic samples depend on the observed data. We propose a bias-correction procedure that estimates the generator-induced loss discrepancy from a held-out subset of majority observations and transfers this correction to the minority class under a uniform bias-transfer condition. We establish finite-sample bounds for bias transfer and for the excess balanced risk of the resulting empirical risk minimizer, and characterize a regime in which SMOTE induces non-negligible loss bias. The framework can also be implemented in imbalanced multi-task learning and propensity-score estimation, with details provided in the Supplementary Material. Real data analyses show that the correction is most useful when synthetic distortion is appreciable.
Pengfei Lyu, Zhengchi Ma, Linjun Zhang +1
Department of Biostatistics & Bioinformatics, Duke University · Department of Electrical & Computer Engineering, Duke University · Department of Statistics, Rutgers University +1
Numerous studies on Boosting attempt to mitigate classification bias caused by class imbalance. However, existing studies have yet to explore the collaborative optimization of imbalanced learning and model training. This constraint hinders further performance improvements. To bridge this gap, this study proposes a collaborative optimization Boosting model of multiclass imbalanced learning. By integrating the density factor and the confidence factor, this model implements a noise-resistant weight update mechanism alongside a dynamic sampling strategy. Rather than functioning as independent components, these modules are tightly integrated to orchestrate weight updates, sample region partitioning, and region-guided sampling. Thus, this study proposes the collaborative optimization of imbalanced learning and model training. Extensive experiments on 40 public imbalanced datasets demonstrate that the proposed model significantly outperforms seven state-of-the-art baselines. The code and datasets for this paper are available at: https://github.com/ChuantaoLi/DARG.
Chuantao Li, Zhi Li, Jiahao Xu +2
School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu 611731, Sichuan, China · School of Mathematics and Computer, Guangdong Ocean University, Zhanjiang 524088, Guangdong, China