Organizations: School of Computer Science and Technology, Dalian University of Technology, Dalian 116024, China · Key Laboratory of Social Computing and Cognitive Intelligence (Dalian University of Technology), Ministry of Education, Dalian 116024, China · School of Engineering and Computer Science, Victoria University of Wellington, PO Box 600, Wellington 6140, New Zealand · College of Computing and Data Science, Nanyang Technological University, Singapore · School of Mechanics and Aerospace Engineering, Dalian University of Technology, Dalian 116024, China · National and Local Joint Engineering Laboratory of Computer Aided Design, Dalian University, Dalian 116622, China
Imbalanced time-series data appears commonly in diverse applications, such as medical diagnosis. When trained on imbalanced data, classifiers tend to exhibit a bias toward the majority class, making it difficult to identify minority-class instances accurately, despite their critical importance. Sampling techniques have proven effective in rebalancing imbalanced data. However, existing interpolation-based oversampling methods typically rely on pairwise interpolation, limiting their ability to capture high-order inter-sample relationships, and their inherent smoothing effects tend to suppress high-frequency components, compromising spectral fidelity. Therefore, we propose a Frequency-aware Hypergraph-driven Oversampling method (FH-Sampling) for imbalanced time-series classification. In FH-Sampling, a hypergraph is first constructed in the original time-series space to capture complex high-order inter-sample relationships, and sampling ratios are adaptively assigned to mitigate the impact of noisy samples. The discrete wavelet transform decomposes each sample into frequency components. Guided by the sample hypergraph, FH-Sampling combines high-frequency components from related samples while retaining the target sample's low-frequency component to help preserve important frequency characteristics. Experiments on 24 imbalanced datasets demonstrate the advantages of FH-Sampling for imbalanced time-series classification. We further apply FH-Sampling to aircraft trajectory generation. Experiments on simulated aircraft trajectories show that FH-Sampling produces trajectories that better match the real distribution, highlighting its potential for practical applications.
Figures & tables
Fig. 1: Comparison of an ordinary graph and a hypergraph. (a) Edges represent pairwise relationships. (b) Hyperedges represent group-wise relationships, with vertex v3 shared by e1 and e2 .
Fig. 2: The overview of FH-Sampling.
BA
F1-Score
G-Mean
AUC
Method
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
None
0.6290
–
–
0.6266
–
–
0.5670
–
–
0.7527
–
–
SMOTE
0.6353
7.188
1.26E-03
0.6268
7.479
2.95E-04
0.6104
6.562
2.95E-04
0.7555
7.396
7.35E-04
B1-SMOTE
0.6401
5.979
2.29E-02
0.6303
6.312
1.64E-02
0.6119
5.729
1.94E-02
0.7582
6.104
1.79E-02
B2-SMOTE
0.6348
7.479
2.88E-03
0.6282
7.479
2.60E-03
0.6080
6.979
3.50E-03
0.7553
7.604
4.68E-03
SVM-SMOTE
0.6354
7.229
6.50E-04
0.6275
7.271
3.73E-04
0.6032
7.312
1.50E-04
0.7559
7.229
1.24E-03
TABLE I: Comparison using the KNN classifier on the 24 datasets
BA
F1-Score
G-Mean
AUC
Method
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
None
0.6007
–
–
0.5912
–
–
0.4796
–
–
0.8349
–
–
SMOTE
0.6163
6.708
3.86E-02
0.6027
6.833
1.81E-01
0.5613
6.271
3.58E-02
0.8434
6.042
6.84E-01
B1-SMOTE
0.6146
7.667
2.85E-02
0.5992
7.417
6.80E-02
0.5552
7.521
1.77E-02
0.8401
6.875
3.60E-01
B2-SMOTE
0.6130
7.500
7.44E-03
0.5961
7.958
6.79E-03
0.5498
7.396
7.44E-03
0.8385
7.375
3.03E-01
SVM-SMOTE
0.6140
7.771
1.38E-02
0.5977
7.396
7.87E-02
0.5328
7.583
6.50E-04
0.8416
6.375
4.91E-01
TABLE II: Comparison using the SVM classifier on the 24 datasets
BA
F1-Score
G-Mean
AUC
Method
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
Avg.Value
Avg.Rank
p -value
None
0.4850
–
–
0.4468
–
–
0.3122
–
–
0.7552
–
–
SMOTE
0.5054
6.500
3.40E-02
0.4732
6.750
8.39E-02
0.4051
5.333
9.70E-01
0.7596
5.583
7.26E-01
B1-SMOTE
0.5197
5.625
3.45E-01
0.4853
5.875
4.06E-01
0.4128
5.521
8.48E-01
0.7641
5.625
7.90E-01
B2-SMOTE
0.5033
7.042
3.22E-04
0.4632
7.792
1.50E-04
0.3968
7.167
2.77E-01
0.7481
7.792
2.25E-03
SVM-SMOTE
0.5121
6.042
2.49E-02
0.4820
5.250
1.97E-01
0.3931
6.354
1.22E-01
0.7606
5.000
4.73E-01
TABLE III: Comparison using the LSTM classifier on the 24 datasets
Fig. 3: Average performance decrease of each ablation variant relative to the complete FH-Sampling method on the 24 datasets. Decreases are measured in percentage points (pp).
Parameter
BA
F1-Score
G-Mean
AUC
σ(k)
0.0110
0.0090
0.0180
0.0073
σ(θ)
0.0072
0.0077
0.0084
0.0057
TABLE IV: Average parameter sensitivity on the 24 datasets
Fig. 4: Runtime distributions of the 14 sampling methods. Diamonds denote means, boxes show quartiles, and inner lines mark medians. Whiskers follow the 1.5×IQR rule; dots are outliers.
Method
MMD2(×10−3)↓
TimeGAN
162.520 ± 2.213
Diffusion-TS
18.307 ± 1.685
TimeVAE
7.945 ± 0.884
FH-Sampling
4.813 ± 0.351
TABLE V: Generation quality on the simulated aircraft trajectory dataset
Fig. 5: Kernel density estimates for real and generated aircraft trajectory data.
Real-world time-series classification tasks often exhibit class imbalance, which can be extremely severe in some applications. To avoid training biased classifiers on imbalanced data, sampling is one of the most popular data pre-processing techniques because of its classifier-agnostic nature. However, due to the complex temporal dependencies in original time-series data and the scarcity of minority-class samples, existing sampling methods, including interpolation-based oversampling methods and deep learning-based generative models, usually suffer from limited generalization and poor diversity when generating new time-series samples. This paper proposes a Frequency-domain representation-guided Multi-tree Genetic Programming-based oversampling approach (FreMGP) to imbalanced time-series classification, where each individual represents a set of synthetic samples for the minority class. A frequency-domain class-discriminative representation module based on contrastive learning is also developed, guiding the evolutionary search toward high-quality synthetic time-series samples. Experiments on imbalanced time-series datasets demonstrate that FreMGP outperforms existing oversampling methods and consistently improves the performance of different classifiers, including both general machine learning and deep learning models.
Wenbin Pei, Yunrong Hao, Zhen Liu +4
School of Computer Science and Technology, Dalian University of Technology, Dalian 116024, China; Key Laboratory of Social Computing and Cognitive Intelligence (Dalian University of Technology), Ministry of Education, Dalian 116024, China · College of Computing and Data Science, Nanyang Technological University, Singapore · School of Mechanics and Aerospace Engineering, Dalian University of Technology, Dalian 116024, China +3
Diffusion models have demonstrated strong performance in time series modeling due to their ability to progressively capture complex data distributions through iterative denoising. However, existing approaches struggle with frequency-sensitive denoising, high-frequency reconstruction and balancing global trends with local dynamics. To address these limitations, we propose HyFAD, a Hybrid time-frequency Diffusion model with Frequency-Aware embedding for time series imputation. Built upon the DDPM paradigm, HyFAD adopts a coupled time-frequency diffusion framework, in which the reverse denoising proceeds sequentially from the time domain to the frequency domain, enabling coarse-to-fine generation. Specifically, the time-domain diffusion process captures low-frequency global trends, while the frequency-domain diffusion process refines high-frequency spectral components. We further introduce a frequency-aware step embedding that exploits the relationship between diffusion steps and spectral components, providing step-dependent spectral guidance and facilitates more accurate band-wise reconstruction. Extensive experiments on multiple benchmark datasets demonstrate that HyFAD achieves state-of-the-art performance. Our source code is available at https://github.com/hongfangao/HyFAD.
Hongfan Gao, Wangmeng Shen, Bin Yang +1
School of Data Science and Engineering East China Normal University
Synthetic medical time-series generation can alleviate data scarcity and support the development of reliable clinical prediction models. However, existing methods mainly focus on matching the overall distribution and temporal dynamics of real data, which does not necessarily ensure strong downstream utility on imbalanced medical datasets. Clinically informative patterns often occur at heterogeneous temporal scales, while rare minority-class characteristics can be obscured by dominant population patterns. To address these challenges, we propose MedFlow, a class-aware multi-scale flow matching framework for medical time-series synthesis. MedFlow employs a vector-quantized multi-scale tokenizer to represent medical sequences at complementary temporal resolutions, capturing both coarse clinical trends and fine-grained dynamics. We further introduce Token Marginal Guidance, which incorporates class-conditional token statistics directly into the flow matching process to steer generation toward class-specific regions of the learned tokens. This mechanism strengthens minority-class patterns, while preserving the global and tail distributions of real data. Experiments on four public datasets covering electronic health records, EEG, and ECG signals demonstrate that MedFlow consistently outperforms recent state-of-the-art diffusion-based baselines across downstream prediction tasks. On average, it improves AUPRC by 5.8%, reduces Context-FID by 88.6%, and achieves 3.8× higher sampling throughput.
Yanhao Huang, Shibo Feng, Wanjin Feng +2
Shanghai Jiao Tong University · Nanyang Technological University · Tsinghua University