feature_importance

[Ehsan86]

I realized the lag and rolling features created in MLForecast are not included in the "feature_importance". The feature importance retrieves only the features transformed in "preprocessor". How can I make sure the lag and rolling features are also included in feature importance?

window_size = [2, 4, 12, 48, 336, 672, 1008] # test_error = 1.258, 1.07

lag_transforms = defaultdict(list)
for lag_ in [1]:
    for i in window_size:
        lag_transforms[int(lag_)].append(RollingMean(window_size=i))
        lag_transforms[int(lag_)].append(RollingMax(window_size=i))
        lag_transforms[int(lag_)].append(RollingMin(window_size=i))


def initialize_pipeline(params):
    pipeline = make_pipeline(preprocessor, RandomForestRegressor(**params))

    forecast_model = MLForecast(
        models={"model": pipeline},
        freq="30T",
        lags=[int(lag) for lag in nominated_lags],
        lag_transforms=dict(lag_transforms)
    )
    return forecast_model
param_grid = {
    "n_estimators": [50],
    "min_samples_split": [10], 
    "min_samples_leaf": [12],
    "n_jobs": [-1],
}
forecast_model = initialize_pipeline(params)
forecast_model.fit(df_train_fold, static_features=["unique_id"], dropna=False, fitted=True)
forecast_model.models_["model"].named_steps["randomforestregressor"].feature_importances_)

[jmoralez]

What does your preprocessor do? Is it letting the lag features go through? You may be dropping them there.