Time Series Generation
Overview
The Time Series Generation project develops state-of-the-art deep generative models for synthesizing high-quality, long-sequence time series data. Our recent contribution, TransFusion, combines the strengths of diffusion models and transformer architectures to overcome limitations in existing time series generation methods, enabling the creation of realistic synthetic temporal data at higher sequence lengths.
The Challenge
Traditional generative models for time series face several limitations:
1. Sequence Length Limitations
- RNN/LSTM-based GANs: Struggle with sequences beyond 100 time steps due to vanishing/exploding gradients
- CNN-based models: Limited receptive fields make capturing long-range dependencies difficult
- Standard GANs: Training instability worsens with longer sequences
2. Temporal Coherence
- Maintaining realistic temporal correlations across long sequences
- Preserving both short-term fluctuations and long-term trends
- Capturing seasonal patterns and periodic behaviors
3. Training Instability
- Mode collapse in GANs becomes more severe with complex temporal patterns
- Difficulty balancing discriminator and generator learning
- Convergence issues with long-sequence modeling
4. Scalability
- Computational cost grows prohibitively with sequence length
- Memory constraints limit batch sizes and model capacity
- Slow generation speed hinders practical applications
Related Publications
Representative Synthetic Data for Fair Decision Making
Md Fahim Sikder
2025
Deep generative models and representation learning techniques have become essential to modern machine learning, enabling both the creation of synthetic data and the extraction of meaningful features …
TransFusion: Generating Long, High-Fidelity Time Series using Diffusion Models with Transformers
Md Fahim Sikder, Resmi Ramachandranpillai, Fredrik Heintz
MLWA – 2025
The generation of high-quality, long-sequenced time-series data is essential due to its wide range of applications. In the past, standalone Recurrent and Convolutional Neural Network-based Generative …
Generative Models for Time-Series and Fair Data Generation
Md Fahim Sikder, Resmi Ramachandranpillai, Daniel de Leng, Fredrik Heintz
SSBA/SSDL 2024 as extended abstract – 2024
Generating Private and Fair Long-Sequenced Longitudinal Healthcare Records
Md Fahim Sikder, Resmi Ramachandranpillai, Fredrik Heintz
SAIS 2023 as extended abstract – 2023
Generating long-sequenced longitudinal healthcare records is critical as it has numerous potential applications. Long-sequenced longitudinal data allow us to better understand and find patterns from …
Synthesizing Longer and Higher-Fidelity Time-Series Data Using Generative Adversarial Networks
Md Fahim Sikder, Resmi Ramachandranpillai, Fredrik Heintz
2023
Sequential Ice Hockey Events Generation using Generative Adversarial Network
Md Fahim Sikder
Student Project Competition, LINHAC 2022 – 2022
We have generated events and coordinates that lead to a hockey goal in this project. Swedish Hockey League data from season 2020-21 for twenty matches provided by Sportlogiq were used that contain …
Privacy-Preserving Synthetic Trajectory Data Generation
David Bergström, Md Fahim Sikder, Fredrik Heintz
SAIS 2020 as extended abstract – 2020
A major open research challenge is developing privacy-preserving machine learning methods that both achieve high performance and privacy guarantees even though the original training data contains …
