Covid-19 Case Prediction
Overview
Trained models to predict covid-19 case trend using traditional ARIMA model and deep learning models including DNN, CNN, RNN, LSTM, cross-validate performance and showcase on web app built by streamlit.
Dash Board
Limited to given storage amount of github (25Mb), only ARIMA and LSTM models are embedded into Streamlit Dash Board
Build ARIMA models
-
Used differentiate method to get stationary data (Generally there are 3 ways to do,Moving average, differentiate, and exponential weighted)
-
Passed Stationary test through Augmented Dickey-Muller statistics (p<0.05) .
-
Chose hyperparameter (p,d,q) through Auto Correlation Function and Partial Correlation Function.
-
Fine-tuned best ARIMA hyperparameters setting by Grid Search
Build Neuron Network models
-
Generated time series tensor data, basically I use past 30 days data to predict cases in next 30 days.
-
A train-in-once model, which generate all countries’ next 30 days prediction by setting output layer neurons to country amount
-
Built connecting layer to deal with time series data when using DNN and CNN model.
-
Hyperparameter tuning.
-
Compared MSE of each models including ARIMA (fine-tuned)
Link
