Documentation for RFSC

Bases: RFSC_base

Implements RFSC algorithm based on parameters inherited from RFSC_base.

__init__(*args)

Inherits parameters from RFSC_base. See help(RFSC_base.init) for more information.

set_attr(params)

Setter for RFSC attributes. Used to update RFSC parameters for DRFSC model.

Parameters:

Name	Type	Description	Default
params	dict	Dictionary of parameters to update.	required

load_data_rfsc(X_train, X_val, Y_train, Y_val, feature_partition=None, sample_partition=None, drfsc_index=None, mu_init=None, M=None)

Loads the data passed from DRFSC into the RFSC object.

Parameters:

Name	Type	Description	Default
X_train	np.ndarray	Training set data	required
X_val	np.ndarray	Validation set data	required
Y_train	np.ndarray	Training set labels	required
Y_val	np.ndarray	Validation set labels	required
feature_partition	list	list of features indices for corresponding drfsc index	None
sample_partition	list	list of sample indices for corresponding drfsc index	None
drfsc_index	tuple	index of DRFSC bin of the form (r,i,j), where r is the drfsc iteration number, i is the vertical partition index, and j is the horizontal partition index	None
mu_init	dict	dictionary containing user-assigned RIPs. If None, RIPs are initialised to 1/n_features.	None
M	dict	dictionary containing relevant previous information for feature sharing	None

rfsc_main(X_train, X_val, Y_train, Y_val, drfsc_index=None)

This is the main section of the RFSC algorithm called by DRFSC. It extracts the model populations and evaluates them on the validation set, and updates the feature inclusion probabilities accordingly.

Parameters:

Name	Type	Description	Default
X_train	np.ndarray	Training set data	required
X_val	np.ndarray	Validation set data	required
Y_train	np.ndarray	Training set labels	required
Y_val	np.ndarray	Validation set labels	required
drfsc_index	tuple, optional	index of DRFSC bin of the form (r,i,j), where r is the drfsc iteration number, i is the vertical partition index, and j is the horizontal partition index	None

generate_models(X_train, Y_train, X_val, Y_val, mu)

Generates random models and for each model evaluates the significance of each feature. Statistically significant features are retained and resultant model is regressed again and its performance on validation partition is evaluated and stored.

Parameters:

Name	Type	Description	Default
X_train	np.ndarray	Training data.	required
Y_train	np.ndarray	Training labels	required
X_val	np.ndarray	Validation data	required
Y_val	np.ndarray	Validation labels	required
mu	np.ndarray	Array of regressor inclusion probabilities of each feature.	required

Returns:

Name	Type	Description
mask_mtx	np.ndarray	Matrix containing 1 in row i at column j if feature j was included in model i, else 0.
performance_vector	np.ndarray	Array containing performance of each model.
size_vector	np.ndarray	Array containing number of features in each model.

update_feature_probability(mask, performance, mu)

Updates the feature probability vector mu based on the performance of the models generated.

Parameters:

Name	Type	Description	Default
mask	np.ndarray	Matrix of shape (n_models, n_features) containing the mask of the models generated.	required
performance	np.ndarray	Performance evaluation for each model.	required
mu	np.ndarray	Current feature probability vector.	required

Returns:

Name	Type	Description
mu_update	np.ndarray	Updated feature probability vector.

tol_check

Checks if maximum difference between mu vectors is below tolerance threshold.

Parameters:

Name	Type	Description	Default
mu_update	np.ndarray	mu at iteration t+1	required
mu	np.ndarray	mu at iteration t	required
tol	float	tolerance condition	required

Returns:

Type	Description
bool	True max difference below tolerance, else False.

select_model

Selects final model based on features that are above the regressor inclusion probability (rip) threshold

Parameters:

Name	Type	Description	Default
mu	np.ndarray	current feature probability vector	required
rip_cutoff	float	regressor inclusion probability threshold	required

Returns:

Name	Type	Description
model_feats	list	list of features that are above the rip threshold

gamma_update

Scale the update of the feature probability vector.

Parameters:

Name	Type	Description	Default
performance	np.ndarray	performance evaluation for each model.	required
tuning	float, optional	tuning parameter to adjust convergence rate, default=10	10

Returns:

Name	Type	Description
gamma	float	Scaling factor for the update of the feature probability vector

generate_model

Takes a vector of probabilities and returns a random model.

Parameters:

Name	Type	Description	Default
mu	np.ndarray	array of probabilities for each feature	required

Returns:

Name	Type	Description
index	np.ndarray	randomly generated numbers corresponding to features ids based on probabilities.

prune_model

Tests whether features are signifincant at selected signifincance level. Returns index of significant features.

Parameters:

Name	Type	Description	Default
model	object	logistic regression model object. See statsmodels.api.Logit	required
feature_ids	list	feature ids included in the model.	required
alpha	float	(0,1) significance level.	required

Returns:

Name	Type	Description
sig_feature_ids	list	list of features above the significance level.