TimeDP

TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts

Papers

• TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts

Overview

Time series generation is a crucial task for many applications, including data augmentation, forecasting, and privacy-preserving data synthesis. Real-world scenarios often require generating time series data that aligns with given examples. TimeDP is designed to address this challenge by learning from multiple domains and generating realistic time series samples based on domain-specific characteristics.

Introduction

Most existing time series generation models focus on single-domain data, limiting their ability to generalize across various real-world scenarios. TimeDP introduces a novel approach that enables multi-domain time series generation using a domain-prompting mechanism. By leveraging diverse time series datasets, TimeDP extracts fundamental patterns and applies them to generate high-quality, domain-specific synthetic time series data.

Key Features

• Multi-Domain Time Series Generation: TimeDP integrates data from various domains and learns shared representations.
• Time Series Basis Extraction: A semantic prototype module constructs a basis library for time series data.
• Few-Shot Domain Prompting: Given a few example time series from a new domain, TimeDP can generate similar time series with high fidelity.
• Diffusion-Based Model: Built upon denoising diffusion probabilistic models (DDPM), ensuring robust and high-quality generation.
• State-of-the-Art Performance: TimeDP surpasses existing baselines in both in-domain and unseen domain generation tasks.

Methodology

1. Input Processing

• Time series data from various domains are fed into TimeDP.
• The model extracts fundamental time series features (prototypes) that serve as elementary building blocks.

2. Learning Time Series Basis

• A semantic prototype module constructs a time series basis library.
• Each prototype represents elementary time series patterns such as trends and seasonality.
• A prototype assignment module (PAM) learns domain-specific prototype weights to construct domain prompts.

3. Generation with Domain Prompts

• During sampling, a small number of example time series from the target domain are provided.
• TimeDP extracts domain prompts and uses them as conditions for generating new time series.
• The generated time series maintains the statistical and temporal characteristics of the target domain.

TimeDP framework overview

Quick Start

Environment Setup

We recommend using conda as environment manager:

conda env create -f environment.yml

Data Preparation

Pre-processed dataset can be found with this link. Set DATA_ROOT environment variable to the path of the dataset, or manually modify the data_path_dict in configs/multi_domain_timedp.yaml to reflect the correct path.

Use main_train.py for model training and visualize.py for domain prompt visualization. The detailed descriptions about command line arguments are as follows:

Parameter Name	Description
base (-b)	Paths to base configuration files.
train (-t)	Boolean flag to enable training. (default: true)
debug (-d)	Boolean flag to enter debug mode. (default: false)
seed (-s)	Seed for initializing random number generators. (default: 23)
logdir (-l)	Directory for logging data. (default: ./logs)
seq_len (-sl)	Sequence length for the model. (default: 24)
uncond (-uc)	Boolean flag for unconditional generation.
use_pam (-up)	Boolean flag to use the prototype assignment module.
batch_size (-bs)	Batch size for training. (default: 128)
num_latents (-nl)	Number of latent variables. (default: 16)
overwrite_learning_rate (-lr)	Learning rate to overwrite the config file. (default: None)
gpus	Comma-separated list of GPU ids to use for training.
ckpt_name	Checkpoint name to resume from for test or visualization. (default: last)

Usage

Training TimeDP:

python main_train.py --base configs/multi_domain_timedp.yaml --gpus 0, --logdir ./logs/ -sl 168 -up -nl 16 --batch_size 128 -lr 0.0001 -s 0

Training without PAM:

python main_train.py --base configs/multi_domain_timedp.yaml --gpus 0, --logdir ./logs/ -sl 168 --batch_size 128 -lr 0.0001 -s 0

Training without domain prompts (unconditional generation model):

python main_train.py --base configs/multi_domain_timedp.yaml --gpus 0, --logdir ./logs/ -sl 168 --batch_size 128 -lr 0.0001 -s 0 --uncond

Visualization of domain prompts:

python visualize.py --base configs/multi_domain_timedp.yaml --gpus 0, --logdir ./logs/ -sl 168 --batch_size 128 -lr 0.0001 -s 0 --uncond

Citation

If you use TimeDP in your research, please cite:

@article{huang2025timedp,
  title={TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts},
  author={Yu-Hao Huang, Chang Xu, Yueying Wu, Wu-Jun Li, Jiang Bian},
  journal={AAAI 2025},
  year={2025}
}