def ridge_ncv_regressor(data,
y,
loss="root_mean_squared_error",
k_outer=5,
k_inner=5,
n_evals=25,
seed=rd.randint(0, 9999),
verbose=True):
"""
conducts ridge l2 regularization for linear regression prediction problems,
returns specified prediction error mean (default: root mean squared error)
designed for rapid prototyping, quickly obtains prediction results by
compromising implementation details and flexibility
applicable only to linear regression problems, unifies three important
supervised learning techniques for structured data:
1) nested k-fold cross validation (minimize bias)
2) bayesian optimization (efficient hyper-parameter tuning)
3) linear regularization (reduce model complexity)
bayesian hyper-parameter optimization is conducted utilizing tree prezen
estimation, linear regularization is conducted utilizing l2
returns custom regression object:
- root mean squared error (or other specified regression metric)
- list of root mean squared errors on outer-folds
data:
- pandas dataframe (n > 2)
- clean (no nan's)
y:
- string
- header of y reponse variable
loss:
- string
- objective function to minimize
- default "root_mean_squared_error"
- supports any error metric found in sklearn
k_outer:
- pos int
- k number of outer folds (1 < k < n)
k_inner:
- pos int
- k number of inner folds (1 < k < n)
n_evals:
- pos int
- number of evals for bayesian optimization
- default 25
seed:
- pos int
- fix to reproduce results
verbose:
- bool
- display function output
"""
## conduct input quality checks
ArgumentQuality(data=data,
y=y,
loss=loss,
k_outer=k_outer,
k_inner=k_inner,
n_evals=n_evals,
seed=seed,
verbose=verbose)
## initiate modeling method
method = Ridge
params = reg_params()
## nested cross-valid bayesian hyper-param optimization
ncv_results = ncv_optimizer(
## main func args
data=data,
y=y,
loss=loss,
k_outer=k_outer,
k_inner=k_inner,
n_evals=n_evals,
seed=seed,
verbose=verbose,
## pred func args
method=method,
params=params)
## regression results object
return ncv_results
def lasso_ncv_regressor(data,
y,
loss="rmse",
k_outer=5,
k_inner=5,
n_evals=1000,
random_state=rd.randint(0, 9999),
standardize=True,
verbose=True):
"""
Conducts LASSO L1 Regularization for Linear Regression prediction problems.
Designed for rapid prototyping. Quickly obtains prediction results by
compromising implementation details and flexibility.
Applicable only to linear regression problems. Unifies three important
supervised learning techniques for structured data:
1) Nested K-Fold Cross Validation (minimize bias)
2) Bayesian Optimization (efficient hyper-parameter tuning)
3) Linear Regularization (reduce model complexity)
Bayesian hyper-parameter optimization is conducted utilizing Tree Prezen
Estimation. Linear Regularization is conducted utilizing L1 shrinkage.
Arguments:
data:
- pandas dataframe (n > 2)
- clean (no nan's)
y:
- string
- header of y reponse variable
loss:
- string
- objective function to minimize
- default "root_mean_squared_error"
- supports any error metric found in sklearn
k_outer:
- pos int
- k number of outer folds (1 < k < n)
k_inner:
- pos int
- k number of inner folds (1 < k < n)
n_evals:
- pos int
- number of evals for bayesian optimization
- default 25
random_state:
- pos int
- fix to reproduce results
standardize:
- bool
- standardizes data
verbose:
- bool
- display function output
"""
## conduct input quality checks
ArgumentQuality(data=data,
y=y,
loss=loss,
k_outer=k_outer,
k_inner=k_inner,
n_evals=n_evals,
random_state=random_state,
standardize=standardize,
verbose=verbose)
## initiate model
method = Lasso
params = reg_params()
## nested cross-valid bayesian optimization
ncv_results = ncv_optimizer(
## main func args
data=data,
y=y,
loss=loss,
k_outer=k_outer,
k_inner=k_inner,
n_evals=n_evals,
random_state=random_state,
standardize=standardize,
verbose=verbose,
## pred func args
method=method,
params=params)
## regression results
return ncv_results