Table 2 Works covered in “Using LSL to solve existing imbalance” section
Proposed model | Application | \(\rho\) | Improvement vs. best baseline | Baseline(s) |
|---|---|---|---|---|
MetaBalance [94] | Image classification | \(\rho = 100\) \(\rho = 1000\) | + 4% Acc. (\(\rho = 100\)) + 6.5% Acc. (\(\rho = 1000\)) | Non-LSL SOTAs + CI Measures |
MetaBalance [94] | Facial recognition | \(\rho\) = 10 | + 2–3% Acc. | Non-LSL model + CI measure |
N/A [95] | Image/video classification | \(\rho \le 1000\) | + 5–9% Acc. + 5% F-Measure | Some LSL SOTAs |
DRAGON [96] | Image classification | \(\rho \approx 250\) | + 1–3% Acc. 10-shot | Four LSL SOTAs |
N/A [97] | Incremental image classification | UNSP | + 25–30% novel-class Acc. | Non-LSL baseline |
RF-MML [98] | Aerial photograph classification | \(10 \le \rho \le 30\) | − 1–2% Acc. on majority + 6–11% Acc. on minority | Non-LSL model + SOTA CI measures |
N/A [99] | Dermatology image classification | \(\rho > 20\) | + 1–2% Acc. | Many SOTAs + CI measures |
FedFew [100] | Federated medical image classification | \(\rho = 10\) | + 20–25% Acc. + 50% precision | Two non-LSL models |
fs-WAE [101] | Bearing fault classification | \(\rho \le 20\) | + 2.5% F-score | Some non-LSL SOTAs |
SSN [102] | Character recognition | UNSP | + 0.4–0.7% Acc. | Some LSL SOTAs |
N/A [103] | Façade defect classification | \(\rho \approx 110\) | + 8–9% overall Acc. + 25–30% novel-class Acc. | One LSL model + proposed variants |
GMDB-fs [104] | Genetic disorder classification | \(\rho \approx 50\) | + 6–10% Acc. | One non-LSL SOTA |
Prototypical Net [105] | Fabric defect classification | \(\rho \approx 75\) | + 4% Acc. | Some non-LSL SOTAs |
DualFusion [106] | Incremental road object detection | \(\rho \approx 1000\) | + 8% AP on novel classes − 3% AP overall | One LSL SOTA |
FSAD-Net [107] | Polyp detection | \(\rho > 100\) | + 7% AUC-ROC | Many LSL SOTAs |
SAPNet [108] | Human-object interaction | UNSP | + 20% Acc. 5-shot + 5% Acc. 1-shot | Some LSL methods |
DSCNN [109] | Hardware Trojan detection | UNSP | + 5–10% Acc. | Many non-LSL SOTAs |
LST [110] | Open-world image segmentation | \(\rho \ge 1000\) | + 10–12% AP on rare classes + 2% AP overall | Proposed variants + CI measures |
SSF-ViT [111] | Facial expression recognition | \(1 \le \rho \le 40\) | + 0–2% Acc. non-LSL + 2% Acc. 5-shot | Many non-LSL SOTAs |
N/A [112] | Streamer action recognition | UNSP | + 20% Acc. | Two SOTA models |
N/A [112] | General action recognition | UNSP | + 1–3% Acc. | Many SOTA models |
N/A [113] | Gestational age estimation | \(\rho \approx 16\) | − 12% Acc. | SOTA non-ML method |
SIHTD [114] | Hyperspectral object detection | N/A | + 1.3% AUC-ROC + 0–30% “Overall” AUC | Some SOTA non- LSL models |
MetaBalance [94] | Credit card fraud detection | \(\rho \approx 600\) | + 0.6% AUC-ROC | Many CI measures |
MetaBalance [94] | Loan default prediction | \(\rho \approx 4\) | + 1.2% AUC-ROC | Many CI measures |
I-SiamIDS [115] | Network intrusion detection | \(\rho \approx 650\) | + 6% F1 Score (majority) + 5–25% F1 Score (minority) | Four non-LSL models |
N/A [116] | Network anomaly detection | \(\rho \approx 250\) | + 0.5–1% Acc. | Three non-LSL models |
SifterJIT [117] | Software defect prediction | \(\rho \approx 10\) | + 9% AUC-ROC + 2.5% F1 Score | One non-LSL SOTA + CI measure |
SAT-GAN [118] | Database error detection | \(\rho \le 30\) | +0-2.5% precision \(\pm 1\%\) F1 Score | Some non-LSL SOTAs |
Meta-IP [119] | Project extension forecasting | \(\rho \le 30\) | + 0.5–2% ROC-AUC + 2–6% BACC | Some CI measures |
MVC [120]\(^{a}\) | 3D point cloud segmentation | \(\rho \approx 40\) | + 0.2% Acc. overall + 2% Acc. on rare classes | Four non-LSL SOTAs |
N/A [121] | Electricity theft detection | \(1.5 \le \rho \le 9\) | + 6% AUC-ROC (\(\rho\) = 9) + 10–12% F1 Score (\(\rho\) = 9) | Six non-LSL models |
N/A [122] | Electrocardiogram classification | UNSP | + 10% Acc. 50-shot + 40% Acc. 3-shot | Three non-LSL models |
MFCCs with FSL [123] | Cough audio diagnosis | \(\rho \approx 15\) | + 1.5% AUC-ROC | Other proposed model + baseline |
N/A [124]\(^{a}\) | Birdsong classification | \(\rho \approx 75\) | + 1.3% Acc. 7-shot + 10.5% Acc. 1-shot | Other proposed models |
N/A [125] | Utterance classification | UNSP | + 4–9% F1 Score | One baseline |
SiameseCHEM [126] | Biochemical activity prediction | UNSP | + 22% Avg. BACC | Three non-LSL models |
Meta-MMFNet [127] | Micro-expression recognition | UNSP | + 5% acc. | Many non-LSL SOTAs |
CBC [128] | Software defect classification | \(5 \le \rho \le 50\) | − 6–12% F1 Score | Two non-LSL rooflines |