def featimp_by_steps(model, steps, sort_step, top, cols):
"""
Return pd.DataFrame of top feature importances by selected steps.
Args:
model: Model to get importances for.
steps: List of ordered steps to include in table.
sort_step: Step to sort table by.
top: Top number of features to include.
cols: Feature list.
"""
for step in steps:
fi_dict = model.extras.feature_importances["test"][step]
step_df = pd.DataFrame(fi.reorder_fi_dict(fi_dict))
step_df = step_df.rename(columns={"importance": f"s={step}"})
step_df.set_index("feature", inplace=True)
df = df.join(step_df) if step > steps[0] else step_df.copy()
df = df.sort_values(by=[f"s={sort_step}"], ascending=False)
df = df[0:top + 1]
return df
def featimp_by_steps(model, period, steps, sort_step, top, cols):
"""
Return pd.DataFrame of top feature importances by all selected steps,
sorted by selected step, from the test partition.
Args:
model: Model to extract feature importances from.
period (str): Period to extract feature importances from.
steps (list): List of ordered steps to include in table.
sort_step (int): Step to sort table by.
top (int): Top number of features to include.
cols (list): Feature list.
"""
for step in steps:
fi_dict = model.extras.feature_importances[period][step]
step_df = pd.DataFrame(fi.reorder_fi_dict(fi_dict))
step_df = step_df.rename(columns={"importance": f"s={step}"})
step_df.set_index("feature", inplace=True)
df = df.join(step_df) if step > steps[0] else step_df.copy()
df = df.sort_values(by=[f"s={sort_step}"], ascending=False)
df = df[0:top + 1]
return df
for model in models:
fi_cm = featimp_by_steps(model=model,
steps=steps,
sort_step=sort_step,
top=top,
cols=model.cols_features)
fi.write_fi_tex(
pd.DataFrame(fi_cm),
os.path.join(out_paths["features"],
f"impurity_imp_{model.name}_{level}.tex"))
sort_step = 3
top = 30
for model in models:
for step in steps:
pi_dict = model.extras.permutation_importances["test"][step]["test"]
step_df = pd.DataFrame(fi.reorder_fi_dict(pi_dict))
step_df = step_df.rename(columns={"importance": f"s={step}"})
step_df.set_index("feature", inplace=True)
pi_df = pi_df.join(step_df) if step > steps[0] else step_df.copy()
pi_df = pi_df.sort_values(by=[f"s={sort_step}"], ascending=False)
pi_df = pi_df[0:top + 1]
fi.write_fi_tex(
pi_df,
os.path.join(out_paths["features"],
f"permutation_imp_{model.name}.tex"))