Time series generation (TSG) is widely used across domains, yet most existing methods assume regular sampling and fixed output resolutions. These assumptions are often violated in practice, where observations are irregular and sparse, while downstream applications require continuous and high-resolution TS.
Although Neural Controlled Differential Equation (NCDE) is promising for modeling irregular TS, it is constrained by a single dynamics function, tightly coupled optimization, and limited ability to adapt learned dynamics to newly generated samples from the generative model.
We propose Diff-MN, a continuous TSG framework that enhances NCDE with a Mixture-of-Experts (MoE) dynamics function and a decoupled architectural design for dynamics-focused training.
To further enable NCDE to generalize to newly generated samples, Diff-MN employs a diffusion model to parameterize the NCDE temporal dynamics parameters (MoE weights), i.e., jointly learn the distribution of TS data and MoE weights. This design allows sample-specific NCDE parameters to be generated for continuous TS generation.
Experiments on ten public and synthetic datasets demonstrate that Diff-MN consistently outperforms strong baselines on both irregular-to-regular and irregular-to-continuous TSG tasks.
Install the environment using the YAML file: ./environment_diffmn.yml:
conda env create -f environment_diffmn.yml --force --no-depsStocks and Energy data are located in ./datasets. Sine, MuJoCo, polynomial datasets are generated and the scripts are included in ./datasets folder.
utils_data.py provides functions for loading data in both regular and irregular settings. In particular, irregular data are preprocessed using the Python class TimeDataset_irregular, which may take some time to run. Once preprocessing is complete, the processed data are saved in the ./datasets directory for future use.
By setting the time series length and missing values within the script, the results in the paper can be reproduced:
Step 1: The initial MoE NeuralCDE can be trained by run_irregular_moencde.py.
Step 2: Parameterizing the MoE weights can be achieved by jointly training the TS samples and their corresponding MoE weights through script run_diffmn_diffsuion.py.
Step 3: Through Step 2, we generate new samples along with their corresponding MoE weights. These weights are then fed into the pretrained MoE Neural CDE to perform continuous time series generation for each new sample. Finally, the refined high-frequency continuous time series are obtained using the script run_irregular_moencde_continues.py, providing richer temporal information and improving the accuracy of downstream tasks.