learning.core¶
Module: learning.core¶
Core module for learning selection
Functions¶
-
selectinf.learning.core.cross_inference(learning_data, nuisance, direction, fit_probability, nref=200, fit_args={}, verbose=False)[source]¶
-
selectinf.learning.core.infer_full_target(algorithm, observed_set, features, observed_sampler, dispersion, fit_probability=<function gbm_fit_sk>, fit_args={'n_estimators': 500}, hypothesis=[0], alpha=0.1, success_params=(1, 1), B=500, learner_klass=<class 'selectinf.learning.learners.mixture_learner'>, learning_npz=None, single=False)[source]¶ Compute a p-value (or pivot) for a target having observed outcome of algorithm(observed_sampler).
- Parameters
algorithm : callable
Selection algorithm that takes a noise source as its only argument.
observed_set : set(int)
The purported value algorithm(observed_sampler), i.e. run with the original seed.
features : [int]
List of the elements of observed_set.
observed_sampler : normal_source
Representation of the data used in the selection procedure.
dispersion : float
Scalar dispersion of the covariance of observed_sampler. In OLS problems this is \(\sigma^2\).
fit_probability : callable
Function to learn a probability model P(Y=1|T) based on [T, Y].
hypothesis : np.float # 1-dimensional targets for now
Hypothesized value of target.
alpha : np.float
Level for 1 - confidence.
B : int
How many queries?
Notes
This function makes the assumption that covariance in observed sampler is the true covariance of S and we are looking for inference about coordinates of the mean of
np.linalg.inv(covariance).dot(S)
this allows us to compute the required observed_target, cross_cov and target_cov.
-
selectinf.learning.core.infer_general_target(observed_outcome, observed_target, target_cov, learner, fit_probability=<function gbm_fit_sk>, fit_args={'n_estimators': 500}, hypothesis=0, alpha=0.1, success_params=(1, 1), B=500, learning_npz=None)[source]¶ Compute a p-value (or pivot) for a target having observed outcome of from algorithm run on the original data.
- Parameters
observed_outcome : object
The purported value observed, i.e. run with the original seed.
target_cov : np.float((1, 1)) # 1 for 1-dimensional targets for now
Covariance of target estimator
learner :
Object that generates perturbed data.
observed_target : np.float # 1-dimensional targets for now
Observed value of target estimator.
fit_probability : callable
Function to learn a probability model P(Y=1|T) based on [T, Y].
hypothesis : np.float # 1-dimensional targets for now
Hypothesized value of target.
alpha : np.float
Level for 1 - confidence.
B : int
How many queries?
-
selectinf.learning.core.infer_set_target(observed_set, features, observed_target, target_cov, learner, fit_probability=<function gbm_fit_sk>, fit_args={'n_estimators': 500}, hypothesis=[0], alpha=0.1, success_params=(1, 1), B=500, learning_npz=None, single=False)[source]¶ Compute a p-value (or pivot) for a target having observed outcome of algorithm on original data.
- Parameters
observed_set : set(int)
The purported value observed when run with the original seed.
features : [int]
List of the elements of observed_set.
observed_target : np.ndarray
Statistic inference is based on.
target_cov : np.ndarray
(Pre-selection) covariance matrix of observed_target.
learner :
Object that generates perturbed data.
fit_probability : callable
Function to learn a probability model P(Y=1|T) based on [T, Y].
hypothesis : np.float # 1-dimensional targets for now
Hypothesized value of target.
alpha : np.float
Level for 1 - confidence.
B : int
How many queries?
Notes
This function makes the assumption that covariance in observed sampler is the true covariance of S and we are looking for inference about coordinates of the mean of
np.linalg.inv(covariance).dot(S)
this allows us to compute the required observed_target, cross_cov and target_cov.