CaMUL: Calibrated and Accurate Multi-view Time-Series Forecasting

Setup

We require you to have anaconda or miniconda installed. Run the script ./scripts/setup.sh to setup the virtual environment with all the required packages.

Data download and preprocessing

Twitter dataset

The probability distributions for each week over all states are available in ./data/tweet_dataset folder as npy files for each week and state.

Power dataset

Run the ./scripts/download_power.sh to download dataset.

Covid dataset

Covid dataset is available in ./data/covid_data folder. Run ./scripts/covid_preprocess.sh to preprocess the features of dataset.

Google Symptoms

Symptoms dataset is available in ./data/symptom_data. Run ./scripts/preprocess_symp.sh for preprocessing.

Experiments

We have ./train_tweets.py, ./train_covid.py, ./train_power.py, ./train_symp.py to run the model for each of the benchmarks. You may tune the arguments related week ahead, prediction week/season by passing the commandline arguments. Use the --help flag for a list of all arguments.

Acknowledgement

In case you use our code or datasets, please cite us as:

@article{kamarthi2021camul,
  title={CAMul: Calibrated and Accurate Multi-view Time-Series Forecasting},
  author={Kamarthi, Harshavardhan and Kong, Lingkai and Rodr{\'\i}guez, Alexander and Zhang, Chao and Prakash, B Aditya},
  journal={Proceedings of the Web Conference 2022},
  year={2022}
}