ReCTSi: Resource-efficient Correlated Time Series Imputation via Decoupled Pattern Learning and Completeness-aware Attentions
Welcome to the repository for the paper "ReCTSi: Resource-efficient Correlated Time Series Imputation via Decoupled Pattern Learning and Completeness-aware Attentions". This repository provides access to the code, and datasets associated with our submission.
This is currently a preliminary release of an informal draft version, which may have minor differences or issues. The official version will be released soon.
To set up the required experimental environment, we recommend using Anaconda. Create and activate an environment with the following commands:
# Create the environment
conda env create -f conda_env.yml
# Activate the environment
conda activate rectsi
We have implemented ReCTSi using several datasets, including traffic datasets (PeMS-BA, PeMS-LA, and PeMS-SD), an air quality dataset (AQ36), and an infection case dataset (COVID-19).
Download the datasets from Google Drive (courtesy of PoGeVon from KDD 2023). After downloading, unzip and move them into the dataset
folder at the root of this repository.
Model | Venue | Year | Link |
---|---|---|---|
BRITS | NeurIPS | 2018 | Link |
rGAIN | AAAI | 2021 | Link |
SAITS | ESWA | 2022 | Link |
TimesNet | ICLR | 2022 | Link |
GRIN | ICLR | 2022 | Link |
NET3 | WWW | 2021 | Link |
PoGeVon | KDD | 2023 | Link |
To run experiments and compute metrics for deep imputation methods, use the run_imputation.py
script. Here`s an example command:
python run_imputation.py
For experiments with the PEMS datasets, adjust the subdataset_name
value in the pems.yaml
configuration file to match the specific dataset (PEMS-04
for PeMS-BA,PEMS-07
for PeMS-LA,PEMS-11
for PeMS-SD).
This code builds upon the implementations of GRIN and PoGeVon. We extend our gratitude to their contributions to the field.