TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching

This repository contains the PyTorch implementation of ARROW, "TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching". In this paper, we propose TiWeaver to adaptively handle temporal dynamics across diverse multivariate time series. Our model include four core modules: Time Series Embedding, Graph-Guided Adaptive Tokenizer (G$^2$AT), Fine-Grained Asynchronous Dependency Extractor (FADE) and Prediction Header.

Install

First, you should install the dependencies as listed in requirements.txt and activate the environment:

conda create -n TiWeaver python=3.10
conda activate TiWeaver

Then, you should install packages:

pip install -r requirements.txt

Prepaer Datasets

You can obtained the well pre-processed datasets from Google Drive. Create a folder named ./dataset and organize it as follows:

dataset/
├── exchange/
│   ├── exchange.csv
│   ├── exchange_freq.csv
│   └── exchange_masked_20pct.csv
├── weather/
│   ├── weather.csv
│   ├── weather_freq.csv
│   └── weather_masked_20pct.csv
├── zafnoo/
│   ├── zafnoo.csv
│   ├── zafnoo_freq.csv
│   └── zafnoo_masked_20pct.csv
├── HumanActivity/
│   ├── preprocessed/
│   └── raw/
├── USHCN/
│   ├── preprocessed/
│   └── raw/
└── P12/
    ├── preprocessed/
    └── raw/

Run Experiments

We provide the experiment scripts under the folder ./Run/scripts. You can run this repository as follows:

sh .Run/scripts/humanactivity.sh

Citation

If you find this repository useful for your research, please consider citing:

@inproceedings{li2026TiWeaver,
  title={TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching},
  author={Li, Zhe and Tian, Jindong and Miao, Hao and Lei, Zhi and Guo, Chenjuan and Yang, Bin},
  booktitle={SIGKDD},
  year={2026}
}