[New] Prediction of ODE parameters with indicators

[lisphilar]

Summary of this new feature

At the current version, we perform scenario analysis, directly changing ODE parameter values of the future phases. However, it is difficult to estimate the impact of changing policy measures on ODE parameter values.

New method to calculate ODE parameter values with user-defined OxCGRT indicators will be useful.

(Optional) Solution

With #280, short-term prediction of ODE parameter values using OxCGRT indicators was included. The next step is that we add a feature to update OxCGRT dataset with user-defined values.

In addition, it is necessary to improve interpretability of prediction (Elastic Net regression) of ODE parameter values using Scenario.fit_predict().

[lisphilar]

To achieve this, a method of Scenario will be necessary to registers future values of extra datasets.

[lisphilar]

In analysis.DataHandler class internally,

• save subsets for the selected country/province with a dataframe, not instances of data cleaning class. (related to [New] backup/restore Scenario object #741)
• add a method to to register future values of extra dataset.

[lisphilar]

future_regressor functionality of AutoTS package can be used.
https://winedarksea.github.io/AutoTS/build/html/source/tutorial.html#simulation-forecasting-1

[lisphilar]

Class for feature engineering, feature selection, extending (add new records on future dates) and easy visualization will be required.
Named MLEngineer as instance.

[lisphilar]

The new class also Includes forecasting with/without regression.

[lisphilar]

Interface could be as follows.

import covsirphy as cs
snr = cs.ODEScenario.auto_build(geo="Japan", model=cs.SIRFModel)
# Prediction of ODE parameters WITHOUT indocators
snr.build_with_template(name="Predicted", template="Baseline")
snr.predict(days=30, name="Predicted", verbose=1)
# Prediction of ODE parameters WITH indocators
data_eng = cs.DataEngineer()
data_eng.download().clean().transform()
subset_df, *_ = data_eng.subset(geo="Japan")
X_all = subset_df.drop(["Population", "Susceptible", "Confirmed", "Infected", "Fatal", "Recovered"], axis=1)
ml = cs.MLEngineer(verbose=1)
pca_model = ml.pca(X=X_all, n_components=0.95)
X_pca = pca_model["PC"].copy()
X_pred = ml.forecast(Y=X_pca, days=30, X=None)
snr.build_with_template(name="Predicted_with_X", template="Baseline")
snr.predict(days=30, name="Predicted_with_X", verbose=1, X=X_pred)
# Compare scenarios
snr.append()
snr.describe()

[lisphilar]

For ducumentation, new notebook will be added to example.

[lisphilar]

With #1157,

• [New] Prediction of ODE parameters with indicators #477: user can specify the seed via MLEngineer(seed) and ODEScenario().predict(seed)