joelongLin / Shared_SSM

Shared State Space Model

README

This is source code of "Shared State Space Model for Background Information Extraction and Time Series Prediction" (accepted by Neurocomputing)

Author: Ruichu Cai, Zhaolong Lin, Wei Chen, Zhifeng Hao

If this work helps you, please cite our work with bibtex below

@article{cai2021shared,
  title={Shared State Space Model for Background Information Extraction and Time Series Prediction},
  author={Cai, Ruichu and Lin, Zhaolong and Chen, Wei and Hao, Zhifeng},
  journal={Neurocomputing},
  year={2021},
  publisher={Elsevier}
}

You can read our paper in this link.

Data Preparing

• Download gold and rename it as GOLD.csv ,then put it under gluonts/lzl_shared_ssm/data_process/raw_data or Use the dataset in the /attachment
• I put the cryptocurrency dataset and air quality dataset(which are clean) under /attachment, move them to the gluonts/lzl_shared_ssm/data_process/raw_data.
• You can Download the Second Experiment Dataset from SIGIR FINIR 2020 And then put them under the same folder as above dataset. Uncompress them and put them like photo below

Needed python package

• Our Model is implemented by python3.6.5 with tensorflow1.14
• There are serveral baseline are implemented by mxnet. This work get a lot of help from Amazon gluonts, you can use pip install gluonts to help you install major of packages

Quick Start

1. use shell script to train models and make prediction :
   • You Can use bin/run_shared_ssm.sh to quickly run our Shared State Space Model. The Hyperparameter are set same range as the paper.
   • Other baseline script are also put in the same folder and named as bin/run_{BASELINE}.sh
2. evaluate prediction of models :
   • you can use python evaluate/acc_result.py to checkout accuracy or you can use python evaluate/rmse_result.py to checkout the rmse
3. debug with right working space
   • you should checkout that every python script run under gluonts/lzl_shared_ssm if you want to debug our code.

Models Source Code

• All Code are put under gluonts/lzl_shared_ssm
• data_process are folder for raw_data, data preprocessing code and preprocessed data.
• evaluate are for evaluation. We can use acc_result.py or rmse_result.py to test output of models. Our Model store output under evaluate/analysis and evaluate/results
• models contain all main result about our model and baseline models. Our model details are put in shared_SSM.py