roeyjob / influenza

Deep Transformer Models for Time Series Forecasting: The Influenza Prevalence Case

日本語はこちら

None official implementation of this paper.

About paper

• The study deals with a time-series forecasting task using Transformer.
• Experiments were conducted using influenza epidemic data provided by the US CDC.
• Predicted epidemics one week ahead with better accuracy than existing methods such as LSTM, RNN, ARIMA, etc.

usege

predict data

1. get state-by-state influenza case count data from the web site of CDC
2. run bash recipes/set_data.sh path_to_downloaded_data
3. Data of each state will be as below.

	REAGION	YEAR	WEEK	ili_ratio
...	...	...	...	...
30	oregon	2020	32	0.5
...	...	...	...	...

run

train with state-by-state data

bash recipes/run_train_single_state.sh 0 200 0.0001 mse

See here for details.

train with all states

bash recipes/run_train_multi_state.sh 0 1 200 0.0001 mse

See here for details

result

One step ahead prediction trained with all states.

Reference

• Wu, Neo, et al. "Deep transformer models for time series forecasting: The influenza prevalence case." arXiv preprint arXiv:2001.08317 (2020).